EXPLAINABLE REINFORCEMENT LEARNING FOR TRANSPARENT AUTONOMOUS DECISION SYSTEMS

EXPLAINABLE REINFORCEMENT LEARNING FOR TRANSPARENT AUTONOMOUS DECISION SYSTEMS | Authorea try { document.documentElement.classList.add('js'); } catch (e) { } var _gaq = _gaq || []; _gaq.push(['_setAccount', 'G-8VDV14Y67G']); _gaq.push(['_trackPageview']); (function() { var ga = document.createElement('script'); ga.type = 'text/javascript'; ga.async = true; ga.src = ('https:' == document.location.protocol ? 'https://ssl' : 'http://www') + '.google-analytics.com/ga.js'; var s = document.getElementsByTagName('script')[0]; s.parentNode.insertBefore(ga, s); })(); Skip to main content Preprints Collections Wiley Open Research IET Open Research Ecological Society of Japan All Collections About About Authorea FAQs Contact Us Quick Search anywhere Search for preprint articles, keywords, etc. Search Search ADVANCED SEARCH SCROLL This is a preprint and has not been peer reviewed. Data may be preliminary. 27 February 2026 V1 Latest version Share on EXPLAINABLE REINFORCEMENT LEARNING FOR TRANSPARENT AUTONOMOUS DECISION SYSTEMS Author : Miruthula G 0009-0004-8117-7090 [email protected] Authors Info & Affiliations https://doi.org/10.22541/au.177222571.15877912/v1 162 views 74 downloads Contents Abstract Information & Authors Metrics & Citations View Options References Figures Tables Media Share Abstract Reinforcement Learning (RL) enables autonomous agents to learn optimal decision-making methods through their continuous interactions with changing environments. RL-based systems are widely used in fields including robotics and intelligent transportation and smart grid management and healthcare automation and industrial control systems. Most reinforcement learning models operate as black-box systems which identify actions without providing any explanation about their decision-making process. The lack of transparency results in decreased human trust toward autonomous systems while creating security risks and responsibility issues which impact critical autonomous systems. To solve this problem the paper introduces an Explainable Reinforcement Learning XRL system which improves visibility for autonomous decision-making processes. The proposed framework extends standard reinforcement learning systems through its additional explanation generation module which explains system activities by analysing state importance and reward assessment. The system generates understandable explanations for every automated decision by examining state-action value distributions and assessment of policy confidence levels. The system includes attention-based state highlighting and reward decomposition analysis as visual tools to explain decision-making processes. The system uses transparency-aware evaluation metrics to measure three aspects of system performance which are interpretability, explanation consistency, and user trust. The framework operates in a simulated autonomous navigation environment which enables performance testing through comparisons to standard RL implementations. The XRL model provides equivalent reward results to existing solutions yet it boosts interpretability and decision-making understanding without requiring extra computing resources. The results demonstrate that explainability should become a fundamental component of reinforcement learning development to establish dependable and accountable autonomous systems. The proposed approach establishes a connection between advanced learning systems and decision-making methods which humans can comprehend. The research team plans to create two new research paths which involve extending existing frameworks to deep reinforcement learning systems and testing them in safety-critical environments Itroduction The emergence of Artificial Intelligence (AI) has brought major changes to contemporary technological systems through its ability to create machines which can make intricate decisions without the need for human operators. Reinforcement Learning (RL) stands out among different AI approaches because it serves as an effective learning method which enables autonomous systems to learn how to navigate dynamic situations while achieving their long-term goals. Active use of RL-based systems can be found in various fields which include self-driving cars and robot handling systems and intelligent traffic management and medical procedure automation and financial market operations and smart resource distribution. The systems adapt their behaviour based on environmental input and enhancement learning process uses reward-based learning methods to boost system performance. The majority of reinforcement learning models achieve impressive results yet they function as unexplainable black-box systems. Deep reinforcement learning systems use neural networks which transform high-dimensional state data into output actions while failing to explain how they make decisions. The models can achieve high accuracy and efficiency but they lack transparent operations and understandable results. Safety-critical fields like autonomous driving and healthcare robotics and industrial automation face significant challenges because of this hidden system behaviour. The system developers and operators and regulators and end-users need to understand the reasons behind autonomous system decisions in order to establish reliability and safety and accountability. The lack of interpretability in reinforcement learning systems creates various problems for users. First, it limits user trust, as humans are more likely to rely on systems whose decisions they can understand. Second, it complicates debugging and system validation, making it difficult to diagnose unexpected or unsafe behaviour. Third, regulatory frameworks for AI increasingly demand transparency and explainability, especially in domains where decisions directly affect human lives. Therefore, there is a growing need to integrate explainability mechanisms into reinforcement learning architectures without compromising performance. Explainable Artificial Intelligence (XAI) has emerged as a research field that develops methods to solve the transparency problems which come from using advanced machine learning systems. XAI techniques attempt to provide human-understandable explanations by identifying influential features, visualizing attention mechanisms, and generating interpretable representations of model decisions. The field of explainable supervised learning models has achieved important advancements, but researchers still need to study how to make reinforcement learning systems more understandable. The RL system requires extra complexity because its decision-making process needs current information together with expectations about future rewards and the methods used to optimize policies. Researchers developed an Explainable Reinforcement Learning (XRL) system which helps autonomous decision systems display their decision-making processes. The proposed system uses an explanation generation module which works with standard RL systems to show how policies operate and which rewards and state-action pairs contribute to their functions. The proposed framework extends its evaluation criteria beyond reward maximization which defines traditional RL systems as it requires both performance and interpretability to achieve success. The system establishes trust through its structured textual and visual explanations which describe each autonomous action while it maintains system reliability and accountability. The primary contributions of this work are threefold. The first step requires us to find and study how black-box reinforcement learning system limits autonomous systems which operate in safety-critical environments. The second step involves creating a transparent reinforcement learning system which enables users to learn policies while providing an integrated explanation system. The third step introduces new evaluation metrics which measure system transparency to evaluate both interpretability and performance metrics. The simulation-based experiments show that our framework achieves reward optimization results which match existing systems while it provides clearer decision-making information. Motivation for Transparency in Autonomous Systems Autonomous systems have reached the point where their operations now function as essential components within critical infrastructure systems which protect human lives while maintaining economic security and system performance. The systems which operate from self-driving cars to industrial robots and healthcare monitoring systems and intelligent traffic control systems depend on data-driven methods for their operational decision processes. More automated systems require less involvement from people. The decision-making process needs transparency which has become a requirement and not an optional feature. The main reason people seek transparency exists to build trust between parties. People tend to trust and use autonomous systems more when they comprehend how the system makes decisions. People who use an autonomous vehicle expect a detailed account when the vehicle performs an unexpected lane change or emergency brake. Users lose trust in a system because its behaviour remains hidden from them. The process of creating understandable information to display how algorithms function will use transparent methods. Another key motivation is safety assurance. Autonomous systems operate in unpredictable environmental conditions which change continuously. In safety-critical domains such as healthcare robotics or transportation systems, even a minor incorrect decision can lead to severe consequences. Engineers and regulators use transparent decision-making to confirm whether operations follow logical reasoning and accurate environmental data. System behaviour under abnormal conditions can be validated through explainability which helps with debugging and fault diagnosis. The need for transparency arises from organizations needing to comply with regulations. Worldwide governments and regulatory bodies are creating AI governance frameworks which require organizations to maintain accountability while providing explanations of their operations. Organizations need to provide explanations for their automated decisions in most applications which involve public safety. Organizations need to use transparency mechanisms because black-box reinforcement learning models do not meet their requirements for responsible deployment. The development of ethical artificial intelligence requires transparency to create better results. The autonomous systems need to follow ethical standards which society expects from them. A system will create negative results when it chooses to operate through efficiency instead of maintaining safety and fairness. The development of explainable models enables developers to examine their reward systems and policy biases while determining their decision-making approach to create an ethical system operation. The technical aspect of transparency brings benefits to system performance because it produces better outcomes and increases overall system strength. The purpose of reinforcement learning agents is to achieve maximum total rewards throughout an extended period, but the system encounters challenges when the rewards become incorrectly specified or the environment presents unpredictable situations. The process of identifying policy inconsistencies through reward analysis and pathway evaluation helps developers to improve their learning methods through decreasing hidden biases. The process of transparency operates as a mechanism which organizations utilize to enhance their operations. The design of user-centric AI systems establishes a requirement for humans and intelligent systems to work together. The design of transparent autonomous systems enables human operators to monitor system performance while retaining control to make necessary interventions. Users who want to understand automated decisions can do so through decision explanations while they keep track of current events. The multiple factors that drive transparency in autonomous systems include trust building safety protection regulatory needs ethical compliance system reliability and human-AI collaboration requirements. The use of reinforcement learning as the main method for training autonomous agents requires developers to implement explainability features that will make these systems more intelligent and easier to comprehend. Fundamentals of Reinforcement Learning The learning method of reinforcement learning (RL) enables intelligent agents to develop decision-making skills through their interactions with various environments. The RL system employs its learning mechanism which depends on agent testing various actions to see what results occur while they receive reward-based feedback. The agent investigates various possible actions while it learns what results each action brings and receives reward and penalty information. The agent learns to choose actions that yield improved results throughout extended periods by observing which actions lead to better results. Reinforcement learning (RL) employs four essential elements which include agent, environment, state and action as its fundamental structure. The agent represents the decision-making system, while the environment represents the world in which the agent operates. The agent tracks the environment’s present condition at each moment and decides which action to take. The environment creates a new state while it delivers a reward signal to the system. The basic structure of reinforcement learning consists of ongoing interactions between the two systems. The policy functions as the main element which determines how the agent will execute its tasks in reinforcement learning. The policy determines which action should be selected for a given state. The agent needs to develop a policy which enables him to achieve maximum reward throughout all future time periods. The agent needs to investigate new actions while he uses his knowledge about successful actions to make decisions about what to do next. The Markov Decision Process (MDP) framework serves as the standard method for modeling reinforcement learning challenges. The MDP framework provides mathematical representation of agent-environment interactions which include states, actions, transition probabilities, rewards, and a discount factor that distinguishes between present and subsequent rewards. The structured representation enables RL algorithms to acquire optimal decision-making strategies for tasks that involve making sequential decisions. Modern reinforcement learning used deep neural networks which created Deep Reinforcement Learning (DRL). DRL enables agents to manage high-dimensional input which includes images and sensor signals and complex environmental data. The implementation of deep reinforcement learning improves advanced task performance but it simultaneously makes systems more complicated and difficult to understand. The hidden operations of reinforcement learning systems require explainability systems which drive the research work presented in this study. Challenges of Black-Box Decision Making The research you conduct uses information which exists until the month of October in the year 2023. The design of autonomous systems needs deep neural networks together with reinforcement learning and their advanced policy optimization methods. The system functions as a black-box solution because its models succeed in dynamic environments yet their internal workings remain hidden from users. The process of neural computations through multiple layers creates a complex system which prevents people from understanding how machines arrive at their conclusions. The first obstacle which black-box decision-making creates for organizations involves their complete lack of openness. In real-world autonomous applications, such as intelligent transportation or robotic systems, unexpected decisions can have serious consequences. Users and developers need a comprehensive explanation when a system executes an action which seems to be out of the ordinary and involves possible dangers. The system requires its decisions to be understandable because otherwise it will not be possible to determine whether they result from logical reasoning or from model errors and uncertainties. The trust people have in autonomous systems has become a significant problem that needs to be addressed. AI systems require trust to achieve both user acceptance and their implementation across all industries. Users will consider a system to be dangerous when they do not understand its operational patterns. AI systems face adoption difficulties because their safety-related technologies require clear understanding of all system components. The process of debugging and performance analysis becomes more challenging when systems operate with hidden functions. Reinforcement learning agents acquire knowledge through reward signals, but any errors in reward system design or training data will result in unexpected agent operations. Decision-making processes that lack understanding make it difficult to find the source of dangerous or incorrect decisions. Engineers need to test various changes through experimentation because they do not have access to complete reasoning methods, which results in more complicated development processes. The issue becomes worse because of both regulatory requirements and ethical standards that need to be followed. Modern AI governance frameworks require organizations to demonstrate three key attributes transparency and fairness and accountability to their stakeholders. Organizations that develop autonomous systems which lack justification mechanisms for their operational decisions will create systems that fail to meet required compliance standards. Healthcare and transportation sectors need explainability as a technical feature which functions as their regulatory requirement for responsible AI deployment. Human-AI collaboration experiences restrictions because black-box models create obstacles to effective working relationships. Human operators will continue working with autonomous systems because these systems need to operate together with people. Human operators need to comprehend how the system works while keeping track of what happens in their environment to achieve successful teamwork. When decisions are opaque, human oversight becomes weaker, reducing the overall reliability of the system. The explanation of reinforcement learning architectures needs to include explainability because of these combined challenges. The technical and regulatory challenges autonomous systems face today become more complex because their current ability to scale creates additional challenges for black-box decision-making processes. The deployment of reinforcement learning agents into complex and large-scale systems results in greater negative impacts when their decisions remain unexplainable. A single misinterpretation of environmental signals can propagate through sequential decision steps to create cumulative errors. The system lacks interpretability which makes it impossible to forecast long-term behaviour patterns and identify system weaknesses. The need to achieve transparency exists as a basic requirement which must be fulfilled to develop autonomous systems that can operate effectively at scale while meeting social acceptance standards. Overview of Explainable Artificial Intelligence The research field of Explainable Artificial Intelligence (XAI) studies ways to create artificial intelligence systems which people can easily understand and see their complete workings. The decision-making processes of AI models become hard to interpret when deep learning architectures create greater model complexity. The systems produce accurate predictions yet function as unidentified systems. XAI provides solutions to this problem by showing users the methods which models use to transform input data into their output results. The design of traditional machine learning methods like decision trees and rule-based systems and linear regression models already includes built-in model interpretability. The model enables users to see which input variables affect the outcome. The structure of modern deep neural networks includes several hidden layers and nonlinear elements which create almost complete barriers to understanding their prediction methods. The need for systematic explainability methods has grown as a result of this complex situation. Explainable AI methods are divided into two main categories which include intrinsic interpretability and post-hoc explainability. The intrinsic methods develop transparent models which allow users to examine decision rules through their predefined design. The post-hoc methods enable users to understand complex black-box systems through their application after the completion of model training. The methods include feature importance analysis, attention visualization, saliency mapping, and surrogate model approximation. The techniques seek to determine or display which elements affect how a model produces results. The main goal of XAI is to create explanations which people can understand. The explanation should provide clear reasons for the decision taken while showing which features had major impact. The explanations must be simple enough for non-experts to understand but need to maintain their technical accuracy. The explanations in high-risk domains must enable system auditing and debugging together with validation processes to maintain system reliability. The relationship between explainability and user acceptance depends on trust as an essential requirement. Users who comprehend how AI systems reach decisions will develop higher trust in these systems. The use of transparent models establishes definite outcomes which users need to build their trust in automated systems. The process of system explanation establishes a platform for humans and machines to work together because users can verify system operations. XAI demonstrates its significance through both regulatory requirements and ethical standards. Current global AI governance frameworks mandate that organizations demonstrate their capacity for transparent and accountable automated decision-making procedures. Organizations that implement AI systems need to prove their models function with fairness while avoiding any concealed bias. Explainability tools help organizations identify hidden biases while they maintain ethical standards throughout sensitive industries which include healthcare and finance and transportation. Despite advancements in explainable supervised learning, XAI methods face significant challenges when applied to reinforcement learning. The process of reinforcement learning requires users to make decisions at multiple points while they try to achieve optimal long-term rewards through different state changes. The system needs to explain its operations through two types of explanations - which require explanations to include both current inputs and the future expected rewards together with the acquired learning methods. The research study centers on developing specialized methods which will help to build explainable systems for reinforcement learning. The process of explainability assessments creates a dual benefit which delivers technical transparency and supports both model development and operational system management. Developers need to understand every aspect of AI system decision-making to achieve better results through model design changes and data training improvements and automatic error detection. Explainable AI creates advantages for both end users and researchers and engineers who work to keep systems secure throughout their operational lifespan. The system uses decision pathway patterns which XAI detects to create a process for system debugging and performance improvement and safe operation in unpredictable real-world situations. The growing use of artificial intelligence in autonomous systems will make explainability increasingly important. Future AI systems will need to work with humans in environments which require both parties to understand each other. XAI functions as a basic component which AI developers must include in their work to create responsible artificial intelligence systems. The development of explainability systems for reinforcement learning situations will make intelligent systems better at showing their decision processes and maintaining accountability while meeting human standards. Related Work in Explainable Reinforcement Learning Explainable Reinforcement Learning (XRL) has gained substantial recognition after systems began to use reinforcement learning in actual operational settings. The traditional reinforcement learning models operate with the main goal of achieving maximum total rewards through their environmental interactions. The models show strong results yet their results remain difficult to understand. Researchers have developed methods to create more understandable RL policies which human users can easily comprehend. The Deep Q-Networks (DQN) system created by DeepMind researchers established a fundamental advancement in deep reinforcement learning technology. Their research proved that agents could acquire advanced control strategies using only unprocessed visual data. The deep neural network methods function as black-box systems because they do not reveal the reasoning behind their selection of specific actions in particular states. The need for understanding RL systems through research into their explainability became apparent because of this constraint. The first group of related studies investigates explanation methods that use visualizations to demonstrate their findings. Researchers apply saliency maps and gradient-based attribution techniques and attention systems to determine which environmental features affected the agent’s decision-making process. In gaming environments, heatmaps display the areas that the agent studied before making its next move. The visual explanations provide useful information but they only show the factors that influence immediate decision-making processes without revealing the strategies that will lead to future rewards. The second area of research studies policy distillation together with rule extraction methods. The researchers use simple decision tree models and symbolic rule systems to create interpretable models which they use to approximate complex deep reinforcement learning systems. Researchers use learned policies to create structured decision paths which they use to improve human comprehension. People need to choose between two competing goals which require them to maintain high performance levels while they work to achieve complete system understanding. Researchers have studied reward decomposition methods which improve system transparency through their application. The reward system provides multiple components which users can interpret instead of delivering a total reward. Users can see which objectives led to their specific action through this system. The method proves especially valuable for multi-objective reinforcement learning situations which require agents to find an optimal balance between three different aspects. The human-in-the-loop reinforcement learning approach serves as a vital research area for related work. This method uses human input to enhance the educational process. The system establishes an agent’s behaviour through reinforcement learning from human preferences which enables the system to meet human expectations. Human guidance enables explanations to achieve greater relevance through their connection to actual thought processes which people use to understand reality. The development of explainable reinforcement learning models for decision-making systems used in safety-critical fields such as robotics and autonomous driving enables systems to provide real-time decision explanations. The systems produce structured explanations which include two types of outputs that show essential environmental elements and which current actions will lead to future rewards. The system requires these abilities to establish trust with users and to maintain responsibility when it operates without human supervision. The field of research needs to solve multiple problems which prevent complete systems from achieving total interpretability. The existing methods create explanations that show specific actions but they do not provide users with an overview of the complete policy that has been learned. The research movement progresses toward hybrid systems which use interpretable design methods combined with deep reinforcement learning approaches to achieve both effective system performance and understandable system operation in autonomous decision-making processes. Problem Definition and Research Objectives Reinforcement Learning (RL) serves as a strong method which trains autonomous agents to make decisions based on their interactions with changing environments. The RL agents discover their best policies through the process of maximizing cumulative reward signals which does not require them to learn specific operational rules. The system demonstrates excellent performance because it can learn new things but the system creates big problems which make it hard to understand and see its inner workings. The modern RL systems which use deep neural networks as their foundation operate as intricate black-box systems. The systems handle complex state information which they transform through hidden nonlinear processes to create output actions. The systems show excellent performance in simulations and real-world applications but their internal processes remain hidden from users. The complete absence of visibility creates an essential obstacle which prevents people from establishing trust and holding others accountable. The main issue exists because we need to comprehend the reasons which lead an agent to choose its particular course of action within any given situation. In reinforcement learning, actions are chosen based on estimated long-term future rewards rather than immediate outcomes. The system makes decisions based on three components which include learned value functions and policy networks and reward propagation mechanisms. The internal computations remain hidden because they operate through neural parameters which prevent tracking of the reasoning pathway. Real-time autonomous systems face increased complexity because of their use in self-driving vehicles and robotic manipulators and intelligent drones. The systems experience safety risks and financial losses because incorrect decisions create dangerous situations. Stakeholders including developers and regulators and end-users need clear explanations about the agent’s actions when it shows unexpected behaviour. Traditional RL frameworks do not provide mechanisms to justify decisions in human-understandable terms. Another significant problem exists for the system because its reward function lacks transparency. The reward function design determines how an RL agent will behave in different situations. The system develops dangerous or unexpected operational methods through tiny errors which occur during the development of reward signals. The training process maintains internal reward optimization which creates challenges when trying to track which reward element affected a specific choice. This process restricts the capacity to debug systems while making it harder to create secure reward systems. The existing explanation methods which researchers use for RL systems follow a post-hoc explanation approach. The approach aims to produce explanations of model behaviour after the system has established its operating procedures. The methods offer partial system understanding through feature importance and attention map analysis yet they fail to prove that their results accurately depict the system’s internal decision-making process. The trustworthiness of the explanation which the system provides presents a major problem. Current research methods prove inadequate because they only assess local interpretability. Local explanations clarify why a single action was taken in a particular state. The system, however, does not generate global interpretable results which display all executive functions that the system has acquired through its training process. The evaluation process for autonomous decision systems becomes difficult because global insight about system operations remains absent. The process of understanding a system introduces a direct relationship which affects how well the system performs. The performance of decision tree models, which provide complete interpretability, declines when they operate in complex, high-dimensional situations. High-performing deep reinforcement learning models establish a trade-off between their transparency and their ability to achieve precise outcomes. Researchers continue to investigate methods which will help them achieve better performance while maintaining system interpretability. The existing research gap includes user-centered explanation development. Different stakeholders require different levels of explanation detail. Engineers require technical explanations which describe value functions and gradients whereas end-users only need simple natural language explanations. The existing RL explainability frameworks fail to modify their explanations according to audience needs which restricts their actual usage. The main research gap requires the creation of unified reinforcement learning systems which integrate explainability as a fundamental component of their learning processes. Future systems need to merge transparency with high performance and safety validation and human interpretability through a single system design which functions as their main explanation system. The gap needs resolution because it plays a vital role in creating autonomous decision systems which businesses can trust and people can hold accountable. Proposed Explainable RL Framework The Explainable Reinforcement Learning (XRL) framework which we have created establishes decision-making capabilities that include transparent operation through its built-in explainability features. The framework enables agents to perform actions which their analysts can comprehend through complete operational explanations instead of traditional reinforcement systems that only seek to optimize total rewards. The development team aims to build an automated system which delivers outstanding performance while remaining easy for users to comprehend. The framework consists of five core components: Environment Interface, Reinforcement Learning Agent, Attention-Based Policy Network, Reward Decomposition Module, and Explanation Generator. The environment interface provides state observations to the agent and receives selected actions. The agent maintains its continuous interaction with the environment while it selects new policies based on the rewards which it receives. The system uses a deep policy network with an attention mechanism as its central component. The attention layer shows which state features have the greatest impact on action selection. The system tracks feature-level importance by storing attention weights that it collects during every decision-making process. This process establishes a base for producing understandable explanations that do not disrupt the models learning process. The framework uses reward decomposition to create better reward transparency. The reward function divides its single aggregated reward signal into separate interpretable objectives that include safety task efficiency and rule compliance. The system keeps track of each reward element as a separate entity. The system determines how each element affects the ultimate decision process during decision making. The explanation generator operates in two modes: local explanation and global explanation. Local explanations provide reasoning for individual decisions. The framework produces an explanation for each state-action pair which shows the important features together with the main reward elements and the expected future value assessments. The system enables users to observe which specific actions the agent has executed at that moment. Global explanations, on the other hand, summarize the overall learned policy. A surrogate interpretable model, such as a decision tree, approximates the complex deep policy network. The surrogate model displays how the agent permanently behaves while showing the methods the agent uses to reach its objectives. Global interpretability supports system validation and behavioural auditing. The framework uses a natural language generation layer because it needs to create explanations that people can understand. The system converts its technical outputs which include feature importance scores and reward weights into understandable human sentences. The system states that ”The vehicle slowed down primarily due to obstacle proximity and safety priority” instead of showing actual numerical weights. The system includes a consistency verification mechanism which tests whether the explanations maintain their original explanation quality. The system uses explanation outputs to test its internal network activations because it needs to verify that the explanation produced accurately displays its decision-making process. This system stops incorrect explanations which do not match the actual reasoning of the model. The framework operates at its best performance when it functions in real-time systems. The system produces explanations through a process which needs less computational resources while maintaining full, open access to its decision-making process. The framework enables its use in applications which require immediate response through autonomous systems found in robotics and intelligent transportation systems. The Proposed Explainable RL Framework presents a complete system which integrates learning, decision-making, and explanation processes within its architecture. The framework introduces interpretability into its reinforcement learning process which results in improved system transparency that establishes user trust while enabling secure operation of autonomous decision-making systems. System Architecture Design The Explainable Reinforcement Learning XRL system architecture which has been proposed in this study combines two functions through its design. The architecture enables the reinforcement learning agent to learn optimal actions while creating understandable explanations for all its decision-making processes. The system uses a modular design approach which enables it to achieve three objectives through its scalable components and interpretable elements and its capacity to operate in real time. The architecture consists of five primary layers: Environment Layer, Agent Learning Layer, Explainability Layer, Policy Monitoring Layer, and User Interface Layer. The system requires each layer to execute its designated tasks while maintaining compatibility with all other system components. The system uses a layered structure which keeps learning and explanation processes separate yet enables ongoing system communication between the two functions. The Environment Layer functions as the external system which the agent uses to perform its tasks in the simulation. The system delivers state information to the agent while the agent transmits its chosen actions back to the system. The environment generates reward signals which it uses to assess how the agent performs its tasks. The system sends its rewards to both the learning module and the explainability module which uses the rewards to monitor transparent operations. The Agent Learning Layer contains the core reinforcement learning model. The system consists of three components which include a state encoder and a policy network and a value estimation system. The policy network determines the best action for a given state, while the value function estimates long-term expected rewards. The policy network uses an attention mechanism to find which input features the system needs to use for choosing its actions. The Explainability Layer generates both local and global explanations through its explanation generation process. The system gathers information from three sources which include attention mechanism data and reward decomposition results and value estimation outputs. The system uses this data to build interpretable models which explain the reasons behind particular decisions. The layer generates explanations through internal learning signals which show system output instead of using external estimation methods. The architectural design includes a Reward Decomposition Module which divides total rewards into distinct parts that can be understood as safety and efficiency and performance. The system uses this method to establish which goal produced the strongest impact on every selection. The system saves the divided rewards in an explanation archive which serves both auditing and validation functions. The Policy Monitoring Layer maintains continuous observation of agent activities throughout different time periods. The system uses surrogate interpretable models which include decision trees to create global behavioural insights through policy summarization. This layer enables ongoing assessment together with detection of irregularities and strategic evaluation of acquired learning patterns. The User Interface Layer functions as the communication link which connects system operations to all stakeholders. The system provides explanations through two different language formats which include technical terms and natural language. The system displays feature importance scores together with reward contributions for users who require technical information. The system provides simplified textual explanations which describe decisions in understandable terms for users who lack technical knowledge. The architecture operates its data flow through a closed-loop system design. The agent observes the state, selects an action, receives reward feedback, updates its policy, and simultaneously logs explanation-related information. The system uses parallel processing to generate explanations without impacting learning performance. The System Architecture Design provides an explanation system that operates directly within the reinforcement learning process. The architecture enables transparent and accountable operation of autonomous decision systems through its unified framework which combines learning and monitoring with reward analysis and explanation modules. Policy Learning and Reward Optimization Strategy The Explainable Reinforcement Learning (XRL) framework establishes a policy learning component which enables agents to develop optimal decision-making methods while maintaining system transparency. The observed states of the environment in reinforcement learning create a mapping which determines which actions a policy will choose. The goal of policy learning is to achieve maximum expected cumulative rewards which are generated through on going environmental interactions. The proposed framework uses deep policy-based learning which connects theθ parameterized neural network to the π(a|s) policy representation. The network operates as a state input system which generates an action probability distribution as its output. The agent makes action decisions through either deterministic methods or probabilistic methods which depend on the current operational status of the system. The training process implements exploration strategies such as epsilon-greedy and entropy regularization to establish an equal balance between exploration activities and exploitation activities. The implementation of a value estimation system together with the policy network maintains learning stability. The value function predicts the total future benefits which will result from any specific state. The framework achieves better results through policy gradient methods which work together with value-based learning techniques to decrease variance while increasing convergence speed. The hybrid method improves learning stability which is vital for challenging environments that involve many dimensions. The entire behaviour of reinforcement learning agents depends on reward signals which makes reward optimization a fundamental component of the learning process. The proposed framework uses an extensive reward function design which consists of multiple parts that can be understood separately. The system uses multiple rewards which include safety compliance and efficiency and task completion to achieve its goals. The training process involves tracking each reward element which requires separate tracking through independent weight monitoring. The cumulative reward is calculated as a discounted sum of future rewards, controlled by a discount factor γ. The discount factor determines the weightage assigned to upcoming rewards in comparison to present rewards. A properly tuned discount factor ensures that the agent balances short-term gains with long-term strategic planning. The optimization process updates policy parameters in a direction that increases expected cumulative returns. To improve understanding of the system, the reward optimization strategy uses reward contribution tracking as an explanation tool. The system tracks the effect of each decomposed reward component on policy updates during every training step. This system enables the explanation module to determine which objective influenced a specific decision the most. The researchers used regularization methods to achieve their goal of obtaining balanced learning results. The system requires penalty terms to stop operators from performing dangerous or unstable movements. The system assigns greater importance to safety penalties when it operates in autonomous driving mode because this protects against hazardous driving behaviour. The system prevents optimization from achieving better performance through optimization methods that compromise safety. The framework uses scheduled policy assessments to evaluate both performance metrics and system stability. The evaluation process uses a controlled environment which prevents any exploration noise from affecting the assessment of the learned policy’s actual abilities. The evaluations detect training problems through their ability to identify overfitting and reward misalignment and unintended behaviour patterns. The system combines policy learning with reward optimization through an explainable framework which enables users to see all performance enhancements. The system tracks three elements which include policy changes and reward impacts and long-term return estimates to show how the agent acquires particular skills. The proposed policy learning and reward optimization method obtains three core outcomes which include efficient operations and stable performance and understandable results. The system development process uses this method to create reinforcement learning systems which possess intelligent capabilities and responsible decision-making abilities which users can trust. Decision Explanation Generation Module The Decision Explanation Generation Module is a core component of the proposed Explainable Reinforcement Learning (XRL) framework. The system requires all reinforcement learning agent actions to be supported by explanations which can be understood by humans. The system provides interpretable explanations of its decisions because the module uses actual learning data to create its explanations which describe all system decisions. The module operates in parallel with the policy network. The explanation module collects relevant internal signals, which include attention weights and value estimates and reward contributions and action probabilities, whenever the agent observes a state and selects an action. The signals create base explanations which show the agent’s thinking process in a true manner. The module uses the current state-action pair to produce local explanations. The system determines which input features created the strongest effect on the decision-making process through its use of attention scores and feature importance metrics. The main factors that control autonomous driving behaviour include obstacle distance and lane position and speed limit from driving environment features. The module calculates how each feature impacts the final results. The system uses two methods to assess feature importance because it needs to measure both feature value and reward value contributions. The explanation shows which reward component led to the agent’s action because the agent optimizes its behaviour based on rewards. This analysis shows the fundamental reasons for decision-making processes. The module also considers long-term value estimation. Reinforcement learning decisions use both future reward predictions and current result assessments to make their decisions. The explanation generation process uses future value score predictions to demonstrate how the agent assessed long-term impacts before making its decision. The module improves its usability through technical explanation output translation into both structured formats and natural language formats. The system provides numerical data to technical users which includes feature weight values and reward impact percentage values. For non-technical users, it generates readable statements such as: “The agent reduced speed primarily due to high obstacle proximity and safety priority.” The system includes a global explanation mechanism which functions through its operational module. The decision logs undergo periodic evaluation to identify the established behavioural patterns which they reveal. The module determines the agent’s overall strategic approach by studying the recurring decision patterns which it has identified. This method enables stakeholders to comprehend the extended period of learning activities which go beyond single learning episodes. The system requires fidelity verification because it functions as the main method for validating explanation trustworthiness. The module verifies whether its generated explanations match the internal policy outputs through a cross-checking process. This system prevents the creation of explanations which misrepresent how the model actually makes its decisions. The design process must consider operational efficiency as an essential requirement. The system generates explanations through a process which functions in real time while maintaining low computational demands. The design enables autonomous systems to maintain their operational speed while providing transparent system functions. The Decision Explanation Generation Module functions as the framework’s main system for producing understandable results. The system allows reinforcement learning agents to present their actions through its combination of feature analysis and reward tracking and value estimation and natural language translation. The module establishes trust and accountability and transparency in autonomous systems through its operational functions. reprint, amsmath,amssymb, aps, ]revtex4-2 Trust Evaluation and Transparency Metrics The trust evaluation process functions as an essential element which enables researchers to measure how well Explainable Reinforcement Learning systems perform their intended tasks. Users establish system trust through performance metrics which include accuracy and cumulative reward and convergence rate but these metrics do not measure user system trust. The framework establishes evaluation metrics which measure transparency and trustworthiness to evaluate system interpretability and user confidence. The trust relationship between users and autonomous systems establishes the boundaries which users will put their faith in automated system decision-making. The trust relationship between users and explainable RL systems depends on the system’s ability to produce clear and consistent and dependable explanations. Users are more likely to accept system results when the system delivers both understandable and accurate explanations of its operational decisions. The first transparency measurement shows Explanation Fidelity as its primary metric. The explanation shows only the actual decision-making process of the RL model when it reaches complete fidelity. The explanation should reach complete fidelity because it needs to show all of the agent’s cognitive processes. The evaluation process requires researchers to compare two elements which include explanation outputs and internal policy activations and reward contributions. The second important metric which we need to measure is Explanation Consistency. The method assesses whether identical states will produce matching explanations. User trust increases when the agent shows similar behaviour across different situations and the explanation module delivers consistent reasoning. Users may lose trust through conflicting explanations which continue to show high performance. The assessment of transparency begins with the measurement of Feature Importance Stability. This metric analyses whether dominant input features maintain their logical connection to domain knowledge. For example, autonomous driving systems should give safety-related features such as obstacle distance their highest importance because those features matter more than any less relevant characteristics. The study demonstrates that stable feature prioritization results in better interpretability across different explanations. The evaluation process includes user-centered metrics which acquire measurements through human feedback studies. The participants need to evaluate the explanations according to three criteria which describe how well they can understand the content. The research team used structured questionnaires to measure two user satisfaction metrics which include user satisfaction score and perceived trust index. The technical evaluation process connects with actual user acceptance through this method. The framework measures Decision Justification Accuracy which evaluates whether the explanation identifies the main reward element that affected the decision. The system must show internal reward logs to prove its assertion that safety was the main deciding factor. The explanation needs to match the optimization behaviour which the system uses to produce its results. The Policy Interpretability Score measures global transparency for assessment purposes. This metric evaluates the accuracy of the surrogate model which uses a decision tree approximation to duplicate the original policy network. The global interpretability of the system improves when surrogate predictions show greater similarity to original model outputs. The system assesses two types of metrics which include transparency measurements and reliability assessment through testing its resistance to noise disturbances and measuring the time needed to produce explanations. Explanation latency ensures that generating explanations does not delay real-time decision-making. A transparent system must maintain both interpretability and operational efficiency. The study evaluates system trustworthiness through two assessment methods which analyse system performance together with its transparent operation capabilities. The evaluation framework ensures that system improvements through enhanced interpretability will not harm its learning capabilities. The research aims to achieve three objectives which include maximum total rewards and accurate explanation delivery and building user trust. The combination of trust evaluation together with transparency metrics creates a complete assessment framework which evaluates explainable reinforcement learning systems. The framework established in this study assesses autonomous decision systems based on four elements which include interpretability and consistency and fidelity and user perception. reprint, amsmath,amssymb, aps, ]revtex4-2 Interpretability Score Clarity of explanation Human rating (1–5) Improves user trust Fidelity Match between explanation & model Perturbation analysis Ensures correctness Transparency Index Overall explainability Composite metric Better adoption Decision Consistency Stability across states Variance measure Reliability improvement Experimental Setup and Simulation Environment The experimental setup is designed to evaluate both the performance and explainability of the proposed Explainable Reinforcement Learning (XRL) framework. The primary goal of the experiment is to assess how well the agent acquires optimal policies while producing dependable and understandable assessment results. The evaluation framework assesses four aspects which include learning efficiency and reward optimization and explanation fidelity and computational overhead. The development of a custom simulation environment creates an autonomous decision-making system which accurately simulates real-world conditions. The environment simulates a control system which requires the agent to drive safely while achieving its performance goals. The simulation tracks multiple state variables which include position and velocity and obstacle proximity and safety indicators. The system updates these state parameters at every time interval. The action space is defined as a discrete set of possible decisions which the agent can choose between. In navigation-based environments actions include moving forward and turning left and turning right and accelerating and decelerating the vehicle. The agent selects actions according to the learned policy network while the environment responds through state transitions and reward feedback. The reward function is carefully structured and decomposed into multiple interpretable components. The components of the system include safety reward to avoid collisions efficiency reward to reach goals quickly and stability reward to maintain smooth control. Each reward component is logged separately during training to support explanation generation and transparency analysis. The deep policy network together with its value estimation module serves as the foundation for implementing the reinforcement learning model. The network uses policy gradient optimization for training which requires a fixed learning rate and discount factor. The system conducts training across multiple episodes until it achieves consistent results. The agent uses exploration strategies during training to learn about different state-action patterns. The system assesses explainability through its system which captures attention weights and reward contributions together with its future value predictions at each decision point. The records serve to create explanations which exist both at local and global levels. The system measures explanation latency to verify that transparency methods maintain their effect on real-time functions. The experimental evaluation includes baseline comparison with a standard deep reinforcement learning model that does not include explainability mechanisms. The study evaluates two different systems by measuring their performance through three specific metrics which include cumulative rewards and their ability to reach optimal performance and maintain consistent policies. The study calculates explanation fidelity and consistency metrics to assess the improvements in interpretability of the system. The researchers measured simulation results through multiple independent training runs to establish both reliability and statistical consistency. The researchers examine three different data points which include average reward curves and explanation accuracy scores and trust evaluation metrics. The comprehensive experimental setup tests the proposed framework across multiple performance metrics while assessing its ability to maintain transparent operations and trustworthy results. The simulation environment enables researchers to create a standardized testing environment which they use to evaluate the performance of the Explainable Reinforcement Learning framework. The experimental design evaluates system performance through two separate assessments which measure both advanced automatic operations and system transparency. Results, Analysis, and Comparative Study The experimental results show that the Explainable Reinforcement Learning framework for XRL and its proposed solution reached similar performance levels while making its operations more understandable. The evaluation compares the proposed model with a baseline Deep Reinforcement Learning (DRL) system that does not include explainability mechanisms. The study examines performance and stability aspects together with explanation fidelity and user trust. The baseline DRL model and the proposed XRL framework both reached learning success after completing similar amounts of training episodes. The cumulative reward curves show that the XRL model achieves almost the same level of reward optimization as its competitors. The explanation generation process created a small increase in computational requirements but the system maintained its performance levels within acceptable boundaries. The proposed framework shows stable learning patterns according to convergence analysis results. The integration of reward decomposition did not negatively impact policy optimization. Decomposed reward tracking prevented the development of unstable learning patterns which emerged from misaligned reward components. The combination of these factors resulted in smoother reward distribution patterns which showed less variation throughout the training process. The proposed XRL framework achieved better explanation fidelity scores than post-hoc explanation methods which were tested on the baseline model from a transparency viewpoint. The fidelity metric confirmed that generated explanations accurately reflected internal policy activations and reward contributions. The process demonstrates its effectiveness because it enables explainability to become an integral part of the learning system design. The analysis of local explanations showed that the main state characteristics matched the rules of the domain. Safety-related variables received the highest priority during decisions that required risk assessment. The stable importance of features between matching states proved that explanations maintained their high degree of consistency. User trust development depends on this particular aspect because it serves as a critical factor for establishing faith in our system. The evaluation of global interpretability through surrogate policy modeling demonstrated that the simplified interpretable model achieved a high degree of similarity with the original deep policy network. The Policy Interpretability Score showed that the surrogate model accurately represented the main behavioural patterns of the RL agent while maintaining high accuracy levels. The baseline DRL model and the proposed XRL system demonstrate their respective advantages through their capability to explain results but their computational requirements create different challenges. The baseline model had slightly faster inference time but the system failed to provide transparent results and trustworthy explanations. The proposed system introduced a small processing overhead but delivered structured, faithful, and real-time explanations, making it more suitable for safety-critical autonomous systems. User-centered trust evaluation further confirmed the benefits of the explainable framework. Participants reported higher confidence levels when explanations were provided alongside decisions. The perceived trust index improved significantly compared to the non-explainable baseline system. This demonstrates the practical value of explainability in real-world deployment. The explanation module maintained its operational stability throughout robustness testing which assessed its performance under conditions of noisy state input. The reward contribution analysis showed accurate objective prioritization because the transparency mechanisms proved to be functional under conditions of moderate environmental uncertainty. The results verify that the Explainable Reinforcement Learning framework achieves its goal because it maintains both system performance and system transparency. The study results demonstrate that when researchers incorporate explainability into reinforcement learning systems their trustworthiness and interpretability increase without causing major efficiency losses. The results demonstrate that it is possible to use transparent autonomous decision-making systems in actual operational environments. Conclusion The research developed a complete Explainable Reinforcement Learning framework which solves the essential requirement of making autonomous decision systems transparent. Traditional reinforcement learning models reach their optimal performance through reward-based systems yet they function as black-box systems which create trust issues and accountability problems for users. The proposed framework provides explainability features to the reinforcement learning system which maintains efficient decision-making operations through its transparent decision-making process. The study demonstrated that transparent operations serve as essential requirements for autonomous systems which operate in safety-critical environments. In intelligent transportation and robotics and automated control system applications, researchers must understand the reasons behind agents’ action choices because this knowledge holds equal significance to their actual selected choices. The framework uses attention mechanisms together with reward decomposition methods and policy monitoring components to deliver accurate and structured explanations about both local and global decision-making processes. The Decision Explanation Generation Module transforms internal computation results into outputs which humans can comprehend. The system tracks three elements which include feature importance and reward contributions and long-term value estimations to create complete action explanation which shows its decision-making process. The system improves technical interpretability together with building user trust in its automatic decision-making capabilities. The experimental results showed that explainability integration does not decrease learning performance. The framework achieved strong cumulative rewards together with consistent convergence while it enhanced explanation accuracy and explanation distribution across different test cases. The proposed system demonstrates a balanced relationship between transparency and computational efficiency which deep reinforcement learning models use as their baseline for comparison. The practical worth of the method received confirmation from trust assessment standards. Users displayed higher trust when they received comprehensive explanations for their decisions. The transparency metrics which included explanation fidelity and consistency and policy interpretability showed that the explanations effectively displayed the internal reasoning process of the agent. This shows that explainable reinforcement learning systems can be used in real-world autonomous systems. The framework needs better capabilities to handle its operation in complex and extensive environments, which remains its principal challenge. Future research can explore adaptive explanation mechanisms, multi-agent explainable reinforcement learning, and integration of causal reasoning models to further enhance interpretability. Explanation formats which users can personalize will improve accessibility for various stakeholders. This research presents a complete and organized framework that enables the integration of reinforcement learning with its inherent ability to provide explanations. The proposed framework establishes a pathway which enables intelligent systems to operate independently while maintaining their ability to be understood by human operators. The integration of explainability into reinforcement learning systems will become vital for maintaining secure and ethical development of AI technologies because of their growing presence in daily human activities. Reference 1. R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018. 2. V. Mnih et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533, 2015. 3. D. Silver et al., “Mastering the game of Go with deep neural networks and tree search,” Nature, vol. 529, no. 7587, pp. 484–489, 2016. 4. F. Doshi-Velez and B. Kim, “Towards a rigorous science of interpretable machine learning,” arXiv:1702.08608, 2017. 5. Z. C. Lipton, “The mythos of model interpretability,” Communications of the ACM, vol. 61, no. 10, pp. 36–43, 2018. 6. M. T. Ribeiro, S. Singh, and C. Guestrin, “Why should I trust you? Explaining the predictions of any classifier,” in Proc. ACM SIGKDD, 2016, pp. 1135–1144. 7. S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” in Proc. Advances in Neural Information Processing Systems (NeurIPS), 2017. 8. D. Gunning, “Explainable Artificial Intelligence (XAI),” Defense Advanced Research Projects Agency (DARPA), 2017. 9. P. W. Koh and P. Liang, “Understanding black-box predictions via influence functions,” in Proc. ICML, 2017. 10. C. Molnar , Interpretable Machine Learning, 2nd ed., 2022. 11. A. Madumal, T. Miller, L. Sonenberg, and F. Vetere, “Explainable reinforcement learning through a causal lens,” in Proc. AAAI Conference on Artificial Intelligence, 2020. 12. T. Miller, “Explanation in artificial intelligence: Insights from the social sciences,” Artificial Intelligence, vol. 267, pp. 1–38, 2019. 13. D. Bahdanau, K. Cho, and Y. Bengio, “Neural machine translation by jointly learning to align and translate,” in Proc. ICLR, 2015. 14. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” in Proc. NeurIPS, 2012. 15. J. Schulman et al., “Proximal Policy Optimization Algorithms,” arXiv:1707.06347, 2017. 16. T. Schaul, J. Quan, I. Antonoglou, and D. Silver, “Prioritized experience replay,” in Proc. ICLR, 2016. 17. C. Amershi et al., “Guidelines for human-AI interaction,” in Proc. CHI Conference on Human Factors in Computing Systems, 2019. 18. B. Goodman and S. Flaxman, “European Union regulations on algorithmic decision-making and a right to explanation,” AI Magazine, vol. 38, no. 3, pp. 50–57, 2017. 19. M. Arrieta et al., “Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI,” Information Fusion, vol. 58, pp. 82–115, 2020. 20. D. Silver et al., “Deterministic policy gradient algorithms,” in Proc. ICML , 2014. Information & Authors Information Version history V1 Version 1 27 February 2026 Copyright This work is licensed under a Creative Commons Attribution 4.0 International License Keywords autonomous systems black box framework problem solving reinforcement Authors Affiliations Miruthula G 0009-0004-8117-7090 [email protected] View all articles by this author Metrics & Citations Metrics Article Usage 162 views 74 downloads .FvxKWukQNSOunydq8rnd { width: 100px; } Citations Download citation Miruthula G. EXPLAINABLE REINFORCEMENT LEARNING FOR TRANSPARENT AUTONOMOUS DECISION SYSTEMS. Authorea . 27 February 2026. DOI: https://doi.org/10.22541/au.177222571.15877912/v1 If you have the appropriate software installed, you can download article citation data to the citation manager of your choice. Simply select your manager software from the list below and click Download. For more information or tips please see 'Downloading to a citation manager' in the Help menu . Format Please select one from the list RIS (ProCite, Reference Manager) EndNote BibTex Medlars RefWorks Direct import Tips for downloading citations document.getElementById('citMgrHelpLink').addEventListener('click', function() { popupHelp(this.href); return false; }); $(".js__slcInclude").on("change", function(e){ if ($(this).val() == 'refworks') $('#direct').prop("checked", false); $('#direct').prop("disabled", ($(this).val() == 'refworks')); }); View Options View options PDF View PDF Figures Tables Media Share Share Share article link Copy Link Copied! Copying failed. Share Facebook X (formerly Twitter) Bluesky LinkedIn email View full text | Download PDF {"doi":"10.22541/au.177222571.15877912/v1","type":"Article"} Now Reading: Share Figures Tables Close figure viewer Back to article Figure title goes here Change zoom level Go to figure location within the article Download figure Toggle share panel Toggle share panel Share Toggle information panel Toggle information panel Go to previous graphic Go to next graphic Go to previous table Go to next table All figures All tables View all material View all material xrefBack.goTo xrefBack.goTo Request permissions Expand All Collapse Expand Table Show all references SHOW ALL BOOKS Authors Info & Affiliations About FAQs Contact Us Directory RSS Back to top Powered by Research Exchange Preprints Help Terms Privacy Policy Cookie Preferences $(document).ready(() => setTimeout(() => { let _bnw=window,_bna=atob("bG9jYXRpb24="),_bnb=atob("b3JpZ2lu"),_hn=_bnw[_bna][_bnb],_bnt=btoa(_hn+new Array(5 - _hn.length % 4).join(" ")); $.get("/resource/lodash?t="+_bnt); },4000)); (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9fe04cdc6e8b593a',t:'MTc3OTE2NDMzMw=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();