{"paper_id":"0af30fbe-1b7d-48d9-926d-74a4aed35ca8","body_text":"1 \n \nAI for Fisheries Science: Neural Network Tools for Forecasting, Spatial Standardization, 1 \nand Policy Optimization  2 \n 3 \nMaia S. Kapur1, Grant Adams1, Marcus Lapeyrolerie2, James T. Thorson1 4 \n1 NOAA Fisheries, Alaska Fisheries Science Center, Seattle, WA 98105 5 \n2 University of British Columbia, Department of Forest Resources Management, Vancouver, BC 6 \nV6T 1Z4 7 \n 8 \nCorresponding author: Maia Kapur kapur.mr@gmail.com  9 \n 10 \nData Availability: Code to obtain and/or re-simulate data for and reproduce the case studies is 11 \navailable at https://github.com/mkapur/deep-fish   12 \n 13 \nAbstract 14 \nThe development of Artificial Intelligence (AI) presents novel opportunities for tackling 15 \ncomplex marine resource management challenges. Among AI models, neural networks are a 16 \npowerful class of tools capable of learning nonlocal and lagged patterns from fisheries data as 17 \nwell as approximating nonlinear relationships among multiple latent variables using estimation 18 \nmethods that automatically implement statistical shrinkage. This gives them potential to 19 \neffectively handle data obtained from fisheries populations subject to dynamic environments. We 20 \nhighlight two flexible subclasses and one application of neural networks:  Long Short-Term 21 \nMemory (LSTM) and Convolutional Neural networks (CNNs) and policy discovery via 22 \nReinforcement Learning. LSTMs are designed to handle sequential data by allowing prediction 23 \nfrom past values at both short and long time-lags. CNNs are not explicitly designed to handle 24 \ntemporal information, but can interpolate a spatial latent variable based on its value within a 25 \ngeographic neighborhood, and can be combined with LSTM models for this purpose. This \"Food 26 \nfor Thought\" paper introduces and applies these neural network approaches, both alone and in 27 \ncombination, to demonstrate their potential application for several foundational topics in 28 \nfisheries science: 1) the forecasting of population weight-at-age, 2) the standardization of spatio-29 \ntemporal indices of relative abundance, and 3) the discovery of harvest policies to optimize 30 \ncatches and maintain spawning biomass. Each section provides a simple, simulated example and 31 \ndescribes the tradeoffs – particularly the lack of inferential capability – presented by using neural 32 \nnetworks over traditional approaches for each topic. We then outline medium-term research 33 \nquestions that may clarify, facilitate or qualify the applicability of these tools to fisheries 34 \nmanagement science.  Finally, we discuss how future combinations of these approaches could 35 \nresult in simplified ways to estimate and forecast stock biomass and provide harvest advice.  36 \n 37 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n1 \n \nIntroduction  38 \nMotivation  39 \nScience based fisheries management is currently facing unprecedented challenges due to shifting 40 \nfunding priorities and dynamic and unexpected changes in the marine environment (Patrick and 41 \nLink, 2015). These challenges are co-occurring with an expansion of methodologies designed to 42 \naddress the diverse data types now available for some fisheries, from molecular information 43 \ninforming movement and demographic estimation (Punt et al., 2024) to fine-scale acoustic 44 \nmonitoring data (Griffin et al., 2018). However, fisheries data have traditionally been analyzed 45 \nusing biological models fit statistically to data, which are constrained by, potentially biased, 46 \nmodel assumptions. Artificial intelligence (AI) tools are less constraining in their assumptions 47 \nand have been explored for data collection and preliminary analysis (such as the statistical post-48 \nstratification of fisheries data (Gasper and Cahalan, 2025, p. 202), occurrence records (Morand et 49 \nal., 2024) or automated image detection (Saqib et al., 2024), but the broader potential of AI 50 \nremains underexplored. There are simultaneous efforts to modernize the modelling infrastructure 51 \nused to assess fishery populations (Maunder et al., 2025), presenting an opportune moment to 52 \nrevisit the methodological landscape of fisheries stock assessment.  This paper focuses on neural 53 \nnetwork models as a promising avenue to modernize scientific fisheries management in the 54 \ncoming decades. We provide readers the core concepts of neural network models, in contrast to 55 \nfamiliar concepts from generalized linear modeling and machine learning and share three 56 \ndemonstrative case studies. 57 \nPointwise Regression: A Common Tool in Fisheries Management 58 \nMost existing fisheries management modeling approaches rely on regression and parametric 59 \nprocess based models. These include traditional approaches such as generalized linear models 60 \n(GLMs) through many machine learning (ML) methods, which include boosted regression trees, 61 \nrandom forests, projection pursuit regression, lasso, and Gaussian process models (to name a 62 \nfew; Hastie et al., 2001), and has been used in fisheries science for over a decade (Rubbens et al., 63 \n2023).  ML and GLMs are similar in that they both typically define a pointwise regression, 64 \nwhere each sample is predicted by a vector of associated covariates; these can include complex 65 \nmethods of handling spatio-temporal processes, such as tinyVAST (Thorson et al., 2025).  Due 66 \nto this conceptual similarity between ML and GLMs, there have been many previous analyses in 67 \nfisheries science comparing ML methods with conventional GLMs or hierarchical models (Stock 68 \net al., 2020), or extending regression models to include a Gaussian process component (Sugeno 69 \nand Munch, 2013; Thorson et al., 2014).  These methods have the advantage of being easy to 70 \nimplement and facilitate inference, but are limited by assumptions regarding the statistical 71 \ndistribution of modeled data and are sensitive to mis-specification. 72 \nNeural Networks: an overview of a foundational AI technique 73 \nNeural networks are a class of machine learning models inspired by the structure and function of 74 \nthe mammalian brain. They are constructed using a series of transformations of features and 75 \npenalized model weights that can approximate any function. This architecture offers advantages 76 \nfor handling high-dimensional, nonlinear and spatiotemporal data types common in fisheries; 77 \ntheir value to ecology for this reason has been recognized for nearly three decades (Lek et al., 78 \n1996). In some cases, neural networks can outperform traditional process-based or statistical 79 \nmodels in predicting ecological processes. Alternatively, neural networks can be embedded 80 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n2 \n \nwithin process based models to improve performance (e.g., (Triebe et al., 2021; Wesselkamp et 81 \nal., 2024)).  82 \n 83 \nWe claim in the following that neural networks (NNs) are often more suited than previous ML 84 \nmethods to identify complex features that are present in fisheries analysis because of their ability 85 \nto learn nonlinear and complex patterns. Neural networks penalize complexity differently from 86 \nGLMs by the progressive updating of model weights (Fan et al., 2021), and can optionally set a 87 \nsubset of model weights to zero to prevent overfitting (Srivastava et al., 2014). These approaches 88 \nare known as “implicit” and “explicit” regularization, respectively. Usefully, these methods for 89 \nregularization do not require marginalizing across any coefficients, and are therefore much faster 90 \nthan regularization in Bayesian or empirical-Bayes hierarchical models.  Current NN methods 91 \nhave also seen less use in fisheries science; one aim of the present paper is to promote the 92 \nexploration of these methods. Table 1 provides an explicit comparison between familiar concepts 93 \nin statistical-ecology and equivalent (though not always identical) concepts in neural 94 \nnetworks/artificial intelligence  95 \nPotential applications of Neural Networks in Fisheries 96 \nThere are numerous types of neural network models, and this paper focuses on two for their 97 \npotential suitability to the spatial and/or temporal dynamism common to marine fisheries. 98 \nRecurrent Neural Networks (RNNs) are a subtype of neural network designed to handle 99 \nsequential data such as time series using feedback loops that maintain an internal state or 100 \nmemory of previous inputs in the series. This recurrent structure enables RNNs to capture 101 \ntemporal dependencies and model the evolution of dynamical systems. They are natural 102 \ncandidates for nonlinear autoregressive models, where future values are predicted based on past 103 \nvalues. 104 \n 105 \nHowever, basic RNNs can suffer from the numerical underflow or overflow issues (see Table 1) 106 \nduring training, which makes it difficult for them to learn long-range dependencies in the data. 107 \nLong Short-Term Memory (LSTM) networks are a specialized type of RNN architecture 108 \ndesigned to mitigate these issues. LSTMs introduce a cell state, which acts as a long-term 109 \nmemory, and sub-states known as “gates” that control the flow of information into and out of the 110 \ncell state, allowing the network to selectively remember and forget information over long 111 \nsequences. The value of RNNs like LSTMs lies in their ability to model and predict the behavior 112 \nof partially observed dynamical systems, where not all relevant state variables are directly 113 \nmeasured. By learning the temporal patterns in the observed data, RNNs can make predictions 114 \nabout future states, which is highly valuable in fisheries science for forecasting population 115 \ndynamics, catch, and other time-dependent variables. 116 \n 117 \nConvolutional Neural Networks (CNNs) were developed for the computer vision field as a 118 \ntechnique to facilitate the detection and labeling of images. In the simplest case, this involves a 119 \nmulti-step process of passing (“convolving”) a set of learnable convolutional filters, or kernels, 120 \nto iteratively extract patterns. This process builds a hierarchical numerical representation, often 121 \nreferred to as a feature map or embedding, that captures high-level features at low spatial 122 \nresolution. This allows the CNN to effectively learn spatial dependencies and patterns in the 123 \nimage, which could be a photograph of an animal or a map of observed biomass from a fishery 124 \nindependent survey, presenting an intriguing possibility for the standardization of spatially-125 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n3 \n \nexplicit data. However, the basic CNN framework is not designed to explicitly handle temporal 126 \ndata nor irregular or sparse datasets characteristic of most fisheries surveys or catch time series. 127 \nResearchers have combined CNN and LSTM neural network approaches to produce forecasts of 128 \nspatial processes (e.g., Yang et al., (2025); the authors are aware of a single example wherein 129 \nCNN and LSTM neural network approaches were combined to produce forecasts of probable 130 \ncatches (Agmata and Guðmundsson, 2025), although other studies have used CNN in isolation 131 \n(Morand et al., 2024). Crucially, that example did not conduct an explicit comparison between 132 \nthe proposed CNN+LSTM approach, variations thereof, and currently-used methods for handling 133 \nspatio-temporal data, such as design-based expansion or regression-based standardization tools 134 \nsuch as tinyVAST, which we do in this case study. 135 \nAims and Structure of this Paper 136 \nThis \"Food for Thought\" paper introduces and applies these neural network approaches, both 137 \nalone and in combination, as a forward-looking demonstration for several foundational topics in 138 \nfisheries science: 1) the forecasting of population processes, with size-at-age as an example; 2) 139 \nthe standardization of spatio-temporal indices of relative abundance, and 3) the discovery of 140 \nharvest policies to optimize catches and maintain spawning biomass. For each case study we 141 \ndescribe the methods used to produce the illustrative example, and discuss the relevance for the 142 \nfisheries assessment and management audience. Where possible, we provide direct mapping 143 \nbetween existing concepts in statistical ecology (such as spatio-temporal standardization or 144 \ninterpolation) or fisheries management (such as management strategy evaluation, MSE; for this 145 \nreason, the section on reinforcement learning for management is longer than the others). 146 \nAdvancing these techniques could enhance the adaptability, precision and sustainability of 147 \nscientific fisheries management in a rapidly changing environment. The paper concludes with a 148 \n“Research Roadmap” outlining the most important research questions to be considered in 149 \nevaluating the tractability of these tools for fisheries management in the coming decades.  150 \n 151 \nCase Studies 152 \nThis paper presents three case studies. Each is designed to highlight how NNs may be applied to 153 \na variety of data types and processes central to fisheries management science. We also selected 154 \nthese to represent a spectrum of tractability and impact, ranging from applications that could be 155 \nimplemented today to those that would require completely novel management frameworks 156 \nFigure 1. We emphasize that the case study methods are examples and are not final; more work 157 \nis needed by the fisheries science community to investigate the tradeoffs, specifications, and 158 \nnuances of each application before they are ready for operational use. 159 \n 160 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n1 \n \nCase Study 1: Neural Networks for Stock Demography Forecasting  161 \nEven highly complex stock assessment models typically resort to simplification of observed 162 \ndemographic processes in order to make management decisions. In particular, projections of 163 \nfuture size-at-age form the basis for fisheries management advice, yet scientists will commonly 164 \nuse a historical or recent (e.g. 5-year) mean of observed fish sizes, though size-at-age can change 165 \ndramatically among years (Stawitz et al., 2019). Stock assessment models also rely on estimates 166 \nof historical size-at-age to derive estimates of historical biomass. Process-based models, such as 167 \nthe von-Bertalanffy growth curve, are commonly used for estimating average size-at-age. Given 168 \nthat size-at-age is well-observed for many managed fisheries, is sensitive to stochastic 169 \nenvironmental processes, and underpins catch advice for industrial fisheries management, this 170 \ncase study illustrates how NN can provide flexibility to predict observable processes in fisheries. 171 \n  172 \nWe specifically compared the performance of two NN approaches to several existing methods 173 \nfor deriving size-at-age. The models included a simple average of size-at-age from the terminal 174 \nfive years of data (‘mean-5’), a three parameter von-Bertalanffy curve (‘VBGF’); a three 175 \nparameter von-Bertalanffy curve with IID year univariate random effects on asymptotic size and 176 \ngrowth rate (VBGF-RE’); a von Bertalanffy curve with 3D AR1 year, age, and cohort random 177 \neffects (‘GMRF’; Cheng et al., 2023); a simple 3-layer NN (‘NN’), and a LSTM (‘LSTM’); see 178 \nSupplementary Material S.1. for further details..  179 \n 180 \nTo explore whether the relative performance of the different approaches depend on the temporal 181 \nstructure of the simulation, we fit all six models to two simulated size-at-age datasets with 182 \nparameters based on Bering Sea pollock (Gadus chalcogrammus) from NOAA’s Alaska 183 \nFisheries Science Center. The first simulated dataset simulated data from a three parameter von-184 \nBertalanffy curve representing constant time-invariant growth. The second simulated dataset 185 \nsimulated data from a three parameter von-Bertalanffy curve with time-varying parameters 186 \nfollowing an AR1 and directional trend representing time-varying growth. Model parameters and 187 \nsample sizes used for simulated data-sets were based on sampling effort and life history 188 \nparameters of cod like species. We produced 300 30-year replicates of each dataset. Only 189 \nsimulations in which all models converged were retained.  190 \n 191 \nWe evaluated prediction performance on each dataset by calculating the average root mean 192 \nsquared error (RMSE) across ages from a 10-fold cross validation where random years of data 193 \nwere removed from each fold. One- and two-year projection performance was evaluated using 194 \nfive retrospective forecast peels. This allowed us to compare average RMSE across peels, both 195 \nfor hindcast and projection accuracy and between the time-invariant or time-dependent scenarios.  196 \n 197 \nThis simulation exercise demonstrated that the LSTM most frequently resulted in the lowest 198 \nRMSE for all three prediction periods (hindcast, one and two years forward) for the time-199 \ninvariant growth simulations (Figure 2). When simulated size-at-age varied through time, 200 \naverage RMSE was most frequently lowest for the GMRF model for the hindcast and one-year 201 \nprojection (Figure 2), though the LSTM and mean-5 performed comparably or better for the 202 \ntwo-year projection in this scenario. 203 \n 204 \nThe GMRF is designed to handle temporal variation explicitly and is able to separate observation 205 \nuncertainty from the latent temporal variability; it is possible that with lower observation error 206 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n2 \n \nthe performance of the LSTM method would have been improved. For both size-at-age 207 \nscenarios, we also visualized  how predictive performance varied across ages, which had a less 208 \ndistinct pattern (Supplementary Figures S1 and S2). 209 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n1 \n \nCase Study 2: Standardizing survey data through space and time using a convolutional 210 \nneural network  211 \nSeveral methods have been developed to model or standardize for spatio-temporal processes in 212 \nfisheries data, particularly survey observations. These methods arose from the recognition that 213 \nunderlying population processes and the methods for observing fish populations are subject to 214 \nvariation in space and time, and accounting for this variation is required to ensure unbiased 215 \nestimates of population size or trajectory. The tinyVAST framework (Thorson et al., 2025) 216 \nbuilds upon the Vector Autoregressive Spatio-Temporal (VAST, Thorson (2019)) modeling 217 \napproach that allows users to specify separable Gaussian Markov Random fields and delta-218 \nGLMMs. The latest approach allows the specification of a broader class and flexibility of 219 \nmultivariate models including spatial factor analysis and ARIMA (Jenkins, 2014) structures, 220 \nenabling the representation of simultaneous, lagged and recursive dependencies common to 221 \necological and fishery processes. TinyVAST is similar to sdmTMB (Anderson et al., 2022) in 222 \nthat they both utilize the modern Template Model Builder language (Kristensen et al., 2016), 223 \nincorporate SPDE-based spatial precision matrices, though the former is better suited for 224 \nallowing for multivariate temporal dependencies. Surprisingly, the recognition that neural 225 \nnetwork models are useful for data with strong spatial dependencies (Wikle and Zammit-226 \nMangion, 2023) has not yet led to rigorous investigation of how these techniques compare to 227 \nexisting approaches for standardizing biomass observations into annual indices of fishery 228 \nabundance. 229 \n 230 \nThe simulation study was conducted on a 20 x 20 spatial grid over 12 years, with spatial 231 \ncorrelation modeled through a row-standardized neighborhood matrix (rho  = 0.95) and temporal 232 \ncorrelation through an AR1 process (p = 0.8). The underlying simulated biomass included both 233 \nspatial and temporal random effects; in each year 200 cells (50% of the domain) were randomly 234 \nsampled for observation that arose from a Tweedie distribution (Zainol Mustafa and Nadia, 235 \n2025; rho = 1.5, psi = 0.2). The experiment was replicated 25 times to assess model performance 236 \nand stability.  For each replicate, observed data were passed to two CNN-based models 237 \n(described below) as well as to tinyVAST to a) interpolate continuous spatio-temporal biomass 238 \nestimates and b) calculate annual indices of abundance. A design-based estimator was included 239 \nfor comparison for the annual indices. Design-based expansion is a simple calculation that takes 240 \nthe average observation across sampled cells and multiplies it across the entire domain based on 241 \nthe fraction of observed cells; this precludes the production of continuous surfaces but facilitates 242 \ncomparison of annual indices.   243 \n 244 \nWe included two versions of the CNN approach, which differed in their handling of missing data 245 \nand temporal dependencies. Both approaches relied on the network to learn spatial relationships 246 \nthrough coordinate embeddings and produced continuous surfaces of estimated biomass. The 247 \nsimpler approach, labeled “CNN” in results, took spatial coordinates as input and processed each 248 \nyear separately; only points with observed (sampled) data were included in training. These 249 \nvectorized inputs passed through two dense embedding layers (dimensions 64 and 128, with 250 \nswish activation), reshaped to a 16x4x2 spatial tensor and processed by two 3x3 two-dimensional 251 \nconvolutional layers with ReLU activation.  The second approach, labeled “CNN-LSTM” 252 \nexplicitly handled sparsity in the observed data via a binary masking channel which informed the 253 \nmodel where observations existed in a given year. Temporal information in the CNN-LSTM was 254 \nincorporated via a lag-based approach, whereby predictions at each timestep were informed by 255 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n2 \n \nthe previous three years of data. Both CNN models were implemented in R using the keras3 256 \npackage (Kalinowski et al., 2025), compiled with Adam and a trainable Tweedie loss function 257 \nwith sigmoid power parameter initialized to 1.5 and constrained to (1, 2) specified to support 258 \ncomparison with tinyVAST. The 12-year, 25-repliate experiment to fit the three models took 16 259 \nhours to run on a personal computer. Performance was compared across models by examining 260 \nthe trajectory of standardized indices, and overall RMSE in log-biomass and log-density (across 261 \nall replicates and years). 262 \n 263 \nThis simulation exercise demonstrated that neither of the CNN-based models out-performed the 264 \nexisting tinyVAST approach in terms of log-biomass and log-density, though all approaches 265 \nwere able to produced standardized indices of similar scale to the true biomass (Figure 3). recent 266 \nwork has replicated our finding that CNNs can match, but not surpass, traditional species 267 \ndistribution modeling approaches (Kellenberger et al., 2026). Examination of the interpolated 268 \nmap for a single year and replicate shows that while both CNN-based models were able to 269 \nidentify areas of relatively higher and lower biomass only tinyVAST was able to recover small-270 \nscale regional patterns. This application of CNNs certainly extends beyond the traditional 271 \nimplementation, wherein an intact image (or collection of images for training) would be passed 272 \nto the model. CNNs are known to perform poorly when images are incompletely observed 273 \n(Heinke et al., 2021), and are also known to perform poorly when training data are limited (most 274 \nCNNs are trained on thousands of images, not 12 years of data as in this case) 275 \n 276 \nThere are promising applications of neural networks that could be used for the detection of 277 \nspatial patterns, some of which are in active development (e.g., (Deng et al., 2025)) and would 278 \nlikely require the generation of continuous spatio-temporal loadings from observed data. 279 \nHowever, this case study indicates that tinyVAST is an appropriately precise and efficient 280 \nmethod for the straightforward task of index standardization using sparse data; more 281 \nsophisticated extensions to the CNN approach such as a vision transformer (Dosovitskiy et al., 282 \n2021) could be revisited for well-sampled fishery populations, and/or from interpolated meshes 283 \nderived from such data. 284 \n 285 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n1 \n \nCase Study 3: Setting next year’s catch using reinforcement learning (RL) 286 \nA common problem is fisheries is the selection of management strategies to achieve policy 287 \nobjectives. Management strategy evaluation is a simulation framework that was developed to 288 \nevaluate alternative strategies under uncertainty. Traditional MSE is a closed-loop forward-in-289 \ntime simulation framework used to evaluate the performance of harvest strategies under 290 \nuncertainty that can include an operating model (OM), estimation model (EM) and management 291 \nrule. MSE is inherently sequential and forward-looking, as it mimics real-world decision-making 292 \nover time. Crucially, the mechanism (or ‘policy’) by which future catches are determined is 293 \nnearly always a predefined rule which is not informed or modified by the simulation itself.  This 294 \nmeans it is not tractable to explore all possible policies using traditional MSE. MSE is a time-295 \nconsuming process that requires stakeholder input (Goethel et al., 2019) and the manual 296 \nspecification of alternative states of nature, estimation models, and harvest policies (Punt et al., 297 \n2016). Recent work has highlighted that static reference points like B0 (unfished biomass, the 298 \nfoundation of many management systems) are highly sensitive to model assumptions (Edgar et 299 \nal., 2024) and fundamentally problematic for data-poor or dynamic fisheries to the degree that 300 \nempirical data streams may offer more robust guidance for management in such situations 301 \n(Blamey et al., 2025). These issues justify the exploration of data-driven methods, such as 302 \nreinforcement learning, that can discover policies autonomously. 303 \n 304 \nReinforcement learning is a subfield of machine learning that is dedicated to solving sequential 305 \ndecision-making problems and consists of two primary components: the agent and the 306 \nenvironment. The RL agent is a trainable neural network that interacts repeatedly with the 307 \n‘environment’, a representation of the population dynamics akin to the operating model used in 308 \nManagement Strategy Evaluation (Butterworth, 2007; Punt et al., 2016). During this interactive 309 \nprocess (called “training”) the agent receives a feedback signal known as the “reward1” which it 310 \nseeks to optimize by updating a policy, analogous to a management rule in MSE. A key 311 \ndistinction between MSE and RL is that the policy in RL is updated according to the agent’s 312 \nexperience in the environment, whereas the HCR in MSE remains static across simulations.  313 \n 314 \nThis case study directly compared the performance of a traditional management strategy 315 \nevaluation with a RL-derived policy for the Eastern Bering Sea (EBS) pollock fishery. This 316 \nallowed us to investigate the tractability and relative performance of the RL approach for a 317 \nfishery with complex dynamics (multiple fishing fleets, age-structure) and a high degree of data 318 \navailability (including weight-at-age and compositional data), yet using a simple, singular 319 \nobjective (maximization of annual harvest). This case study therefore extends beyond recent 320 \ncontributions that focus on simplified or simulated stock dynamics (e.g., Montealegre-Mora et al. 321 \n(2025)). Two recruitment scenarios were explored for both approaches: the first utilized a stock 322 \nindependent approach where expected recruitment was defined by random annual deviations 323 \naround a constant value (‘Constant recruitment’), and the second utilized a Ricker stock-324 \nrecruitment relationship, where stock output varied both following random annual deviations and 325 \nas a function of stock size (‘Ricker recruitment’). For EBS pollock, recruitment dynamics are a 326 \nnoted uncertainty for this stock and both parameterizations have been explored for management 327 \nadvice (Ianelli et al., 2024). This allowed us to determine whether commonly-used recruitment 328 \ndynamics impacted relative performance or stability of the RL approach for this fishery. We 329 \n                                                 \n1 For readers seeking a broader introduction to RL in the context of ecology, see Lapeyrolerie et \nal., (2022). \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n2 \n \ncompared future performance of the catch recommendations determined by the MSE and RL 330 \napproaches by examining the median simulation trajectory and uncertainty of stock spawning 331 \nbiomass realized catches during the projection period. For illustration, we also compared the 332 \nharvest policy learned by the RL approach to the survey observations and realized catches from 333 \nthe MSE. Full specifications of this case study can be found in Supplementary Material section 334 \n2. 335 \n 336 \nThe reinforcement learning (RL) agent trained under constant recruitment developed a harvest 337 \npolicy resembling precautionary harvest control rules. It avoided fishing when bottom trawl 338 \nsurvey biomass was below about 10,500 mt and increased harvest sharply to ~4 million mt above 339 \nthat threshold, reaching ~4.25 million mt at higher levels. This produced large year-to-year 340 \nfluctuations—ranging from zero to nearly double historical catches—but consistently maintained 341 \nspawning biomass above that of the management strategy evaluation (MSE) scenario, which had 342 \nstable catches (~1.5 million mt) and biomass (~2.5 million mt).Under the Ricker recruitment 343 \nscenario, the RL agent’s policy was non-linear but smoother than in the constant recruitment 344 \ncase. It maintained nearly constant catches (1.8–3 million mt) for survey observations above 345 \n4,000 t, unexpectedly increased catches around 2,500 t, and stopped fishing below 1,000 t. This 346 \nstrategy stabilized spawning biomass near 1.25 million mt (~21% of unfished levels; 20%B0 is a 347 \ncommon cutoff for fishery closure) while sustaining harvests of ~3.5 million mt—substantially 348 \nhigher than the MSE scenario, which averaged 1–1.5 million mt of catch and 2–2.5 million mt of 349 \nspawning biomass. Despite its unconventional shapes, the RL policy outperformed the MSE in 350 \ncumulative yield while maintaining more stable trajectories under the more complex Ricker 351 \ndynamics. 352 \n 353 \nThese results expand upon findings by Montealegre-Mora (2025) to suggest that the RL 354 \napproach is capable of discovering harvest policies for complex, data-rich fisheries and produce 355 \ncatch and spawning biomass trajectories of similar magnitude to those obtained by MSE under 356 \ntwo distinct recruitment scenarios. That study demonstrated that RL can provide insights into 357 \nHCR design that conventionally used methods in fisheries management are unable to achieve. 358 \nThis is especially intriguing given that the RL did not need to “step through” an updated 359 \nestimation model at each timestep, presenting a potential companion or alternative to the time-360 \nconsuming process of updating stock assessments in resource-limited settings. 361 \n 362 \nFuture work could explore constraints on the policy curve (i.e., forcing recti-linear policies or 363 \nsetting strict upper limits on catches, though these can inhibit agent training). The RL method 364 \ncould also be extended to base the current state of the population on more data types (age 365 \ncomposition data, for instance) or to include representations of observation or process error (e.g., 366 \nvia curriculum learning, see Table 2). Finally, it would be useful to explore comparison of the 367 \nRL method to empirical harvest control rules that set catches based on recent survey 368 \nobservations, particularly in the context of data quality and availability. The RL method 369 \nemployed here assumes that annual survey observations are equally representative of the 370 \npopulation among years and through time; it would be useful to know if there are lower limits to 371 \nthe frequency and precision of survey observations that facilitate RL learning.  372 \n 373 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n3 \n \nSummary of findings from Case Studies 374 \nOur case studies highlight the flexibility that neural networks present in addressing foundational 375 \ntopics in fisheries science, their immediate shortcomings, and directions for future research. For 376 \nthe prediction of weight-at-age, it appears valuable to include LSTM-based approaches as part of 377 \nan analytical pipeline to evaluate forecast and hindcast projection accuracy given the frequency 378 \nwith which they out-performed other methods in terms of RMSE. For some prediction 379 \ncategories, the GMRF method showed similar performance in the presence of temporal variation 380 \nin growth.  For the development of annual indices of abundance from spatio-temporal survey 381 \ndata, or interpolation of biomass in space, neither the CNN nor CNN-LSTM approaches 382 \nappeared to outperform tinyVAST. It is important to recognize that the CNN technique was 383 \ndeveloped for computer-vision applications under the assumption that the entire image is 384 \navailable to the network; the sparse information provided by our simulated survey was 385 \ninsufficient for the CNN to accurately characterize spatial dependencies, is precisely what 386 \ntinyVAST was designed to handle. The RL-to-MSE comparison exercise highlighted promising 387 \npotential for novel policy discovery, with the tradeoff that absent strong guidance, policies might 388 \nbe un-intuitive (hindering stakeholder communication and support) or induce untenably large 389 \nvariation in population trajectories.   390 \n 391 \nFood for Thought: Priority Research Questions for AI in Fisheries Science 392 \nWhen statistical catch-at-age software tools such as Stock Synthesis became widespread, the 393 \nfisheries science community spent nearly two decades examining the impacts of 394 \n‘misspecification’ on model performance, highlighting how un-accounted-for dynamics in spatial 395 \nstructure, fish growth, selectivity patterns, and more could bias the estimation of management 396 \nquantities and subsequent advice. Now the fisheries science community must apply that same 397 \nfocus to uncovering the trade-offs and biases presented by using neural networks in tandem with 398 \nour existing process-based modeling workflows.  399 \n 400 \nHere, we pose several outstanding questions regarding the potential use of neural network 401 \nmodels for fisheries science, which we categorize based on their use in projections, for setting 402 \nquotas, or for fisheries science and research (Table 3).  In particular, fisheries process models 403 \nwere designed to represent the mechanisms that underly population responses to fishing 404 \n(Beverton and Holt, 1957).  By representing these mechanisms, fisheries scientists presumably 405 \nhoped to develop a “structural causal model” (Pearl, 2009), which could then be used to predict 406 \nfishery responses to policy decisions that have not previously been seen (i.e., the simultaneous 407 \nimpact of fishing and climate change).  We therefore recommend specific comparison of existing 408 \nprocess-based and new neural-network models regarding their ability to predict (and quantify 409 \nuncertainty) about fishery responses to previously unobserved policies; fishery scientists may 410 \nneed to expand their vocabulary for describing sources of uncertainty to include those common 411 \nin the AI literature (for example, Allen Akselrud, 2024).  Our first two case studies emphasized 412 \n“predictive performance,” and good predictive performance is no guarantee of good performance 413 \nfor disentangling multiple causes when recommending new policies (Arif and MacNeil, 2022).  414 \nGiven the large literature regarding causal modelling in artificial intelligence research (reviewed 415 \nbriefly in Luo et al., 2020), we are optimistic that future neural network research can develop 416 \nrobust fisheries policies using monitoring data for systems involving multiple drivers.   417 \n 418 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n4 \n \nConclusion  419 \nThis Food-for-thought paper sought to 1) introduce the scientific fishery management 420 \ncommunity to modern techniques in deep learning with neural networks, exploring specific 421 \napplications to foundational fisheries topics. Our case studies highlight that these methods have 422 \npromise to supplement, improve upon, or change the ways we manage industrial fisheries, and 423 \nunderscore that rigorous benchmarking (comparison to existing methods) should be employed as 424 \nthey are further explored and refined.   425 \n  426 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n5 \n \nReferences 427 \n 428 \nAgmata, A., Guðmundsson, S., 2025. Convolutional-LSTM approach for temporal catch 429 \nhotspots (CATCH): an AI-driven model for spatiotemporal forecasting of fisheries catch 430 \nprobability densities. Biol. Methods Protoc. 10, bpaf045. 431 \nhttps://doi.org/10.1093/biomethods/bpaf045 432 \nAllen Akselrud, C.I., 2024. Random forest regression models in ecology: Accounting for messy 433 \nbiological data and producing predictions with uncertainty. Fish. Res. 280, 107161. 434 \nhttps://doi.org/10.1016/j.fishres.2024.107161 435 \nAnderson, S.C., Ward, E.J., English, P.A., Barnett, L.A.K., Thorson, J.T., 2022. sdmTMB: An R 436 \nPackage for Fast, Flexible, and User-Friendly Generalized Linear Mixed Effects Models 437 \nwith Spatial and Spatiotemporal Random Fields. 438 \nhttps://doi.org/10.1101/2022.03.24.485545 439 \nArif, S., MacNeil, M.A., 2022. Predictive models aren’t for causal inference. Ecol. Lett. 25, 1741–440 \n1745. https://doi.org/10.1111/ele.14033 441 \nBeverton, R.J.H., Holt, S.J., 1957. On the dynamics of exploited fish populations. Chapman & 442 \nHall, London. 443 \nBlamey, L.K., Plagányi, É.E., Moore, T., 2025. Exploring dynamic reference points for a short-444 \nlived, highly variable species under future climate. Can. J. Fish. Aquat. Sci. 82, 1–16. 445 \nhttps://doi.org/10.1139/cjfas-2025-0070 446 \nButterworth, D.S., 2007. Why a management procedure approach? Some positives and 447 \nnegatives. ICES J. Mar. Sci. 64, 613–617. https://doi.org/10.1093/icesjms/fsm003 448 \nDeng, R., Li, Z., Wang, M., 2025. GeoAggregator: An Efficient Transformer Model for Geo-449 \nSpatial Tabular Data, in: The Thirty-Ninth AAAI Conference on Artificial Intelligence 450 \n(AAAI-25). Presented at the Association for the Advancement of Artificial Intelligence, 451 \nAssociation for the Advancement of Artificial Intelligence, pp. 11572–11580. 452 \nDosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, 453 \nM., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N., 2021. An Image is 454 \nWorth 16x16 Words: Transformers for Image Recognition at Scale. 455 \nhttps://doi.org/10.48550/arXiv.2010.11929 456 \nEdgar, G.J., Bates, A.E., Krueck, N.C., Baker, S.C., Stuart-Smith, R.D., Brown, C.J., 2024. 457 \nStock assessment models overstate sustainability of the world’s fisheries. Science 385, 458 \n860–865. https://doi.org/10.1126/science.adl6282 459 \nFan, J., Ma, C., Zhong, Y., 2021. A Selective Overview of Deep Learning. Stat. Sci. 36, 264–460 \n290. https://doi.org/10.1214/20-STS783 461 \nGasper, J., Cahalan, J., 2025. Utilizing the random forest algorithm and interpretable machine 462 \nlearning to inform post-stratification of commercial fisheries data. Fish. Res. 281, 463 \n107253. https://doi.org/10.1016/j.fishres.2024.107253 464 \nGoethel, D.R., Lucey, S.M., Berger, A.M., Gaichas, S.K., Karp, M.A., Lynch, P.D., Walter, J.F., 465 \nDeroba, J.J., Miller, S., Wilberg, M.J., 2019. Closing the feedback loop: on stakeholder 466 \nparticipation in management strategy evaluation. Press. Addressing Fish. Chall. Manag. 467 \nStrategy Eval. 01, 1895–1913. https://doi.org/10.1139/cjfas-2018-0162@cjfas-468 \nmse.issue01 469 \nGriffin, L.P., Brownscombe, J.W., Adams, A.J., Boucek, R.E., Finn, J.T., Heithaus, M.R., 470 \nRehage, J.S., Cooke, S.J., Danylchuk, A.J., 2018. Keeping up with the Silver King: 471 \nUsing cooperative acoustic telemetry networks to quantify the movements of Atlantic 472 \ntarpon (Megalops atlanticus) in the coastal waters of the southeastern United States. 473 \nFish. Res. 205, 65–76. https://doi.org/10.1016/j.fishres.2018.04.008 474 \nHastie, T., Tibshirani, R., Friedman, J., 2001. The Elements of Statistical Learning, Springer 475 \nSeries in Statistics. Springer New York Inc., New York, NY, USA. 476 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n6 \n \nHeinke, D., Wachman, P., van Zoest, W., Leek, E.C., 2021. A failure to learn object shape 477 \ngeometry: Implications for convolutional neural networks as plausible models of 478 \nbiological vision. Vision Res. 189, 81–92. https://doi.org/10.1016/j.visres.2021.09.004 479 \nJenkins, G.M., 2014. Autoregressive–Integrated Moving Average (ARIMA) Models, in: Wiley 480 \nStatsRef: Statistics Reference Online. John Wiley & Sons, Ltd. 481 \nhttps://doi.org/10.1002/9781118445112.stat03472 482 \nKalinowski, T., Allaire, J.J., Chollet, F., 2025. keras3: R Interface to “Keras.” 483 \nKellenberger, B., Winner, K., Jetz, W., 2026. The Performance and Potential of Deep Learning 484 \nfor Predicting Species Distributions. Glob. Ecol. Biogeogr. 35, e70184. 485 \nhttps://doi.org/10.1111/geb.70184 486 \nKristensen, K., Nielsen, A., Berg, C.W., Skaug, H., Bell, B.M., 2016. TMB: Automatic 487 \ndifferentiation and laplace approximation. J. Stat. Softw. 488 \nhttps://doi.org/10.18637/jss.v070.i05 489 \nLapeyrolerie, M., Chapman, M.S., Norman, K.E.A., Boettiger, C., 2022. Deep reinforcement 490 \nlearning for conservation decisions. Methods Ecol. Evol. 13, 2649–2662. 491 \nhttps://doi.org/10.1111/2041-210X.13954 492 \nLek, S., Delacoste, M., Baran, P., Dimopoulos, I., Lauga, J., Aulagnier, S., 1996. Application of 493 \nneural networks to modelling nonlinear relationships in ecology. Ecol. Model. 90, 39–52. 494 \nhttps://doi.org/10.1016/0304-3800(95)00142-5 495 \nLuo, Y., Peng, J., Ma, J., 2020. When causal inference meets deep learning. Nat. Mach. Intell. 496 \n2, 426–427. https://doi.org/10.1038/s42256-020-0218-x 497 \nMaunder, M.N., Crone, P.R., Semmens, B.X., Valero, J.L., Waterhouse, L., Methot, R.D., Punt, 498 \nA.E., 2025. The Center for the Advancement of Population Assessment Methodology 499 \n(CAPAM): A perspective on the first 10 years. Fish. Res. 281, 107162. 500 \nhttps://doi.org/10.1016/j.fishres.2024.107162 501 \nMontealegre-Mora, F., Boettiger, C., Walters, C.J., Cahill, C.L., 2025. Using Machine Learning 502 \nto Inform Harvest Control Rule Design in Complex Fishery Settings. Fish Fish. 26, 1004–503 \n1020. https://doi.org/10.1111/faf.70013 504 \nMorand, G., Joly, A., Rouyer, T., Lorieul, T., Barde, J., 2024. Predicting species distributions in 505 \nthe open ocean with convolutional neural networks. Peer Community J. 4. 506 \nhttps://doi.org/10.24072/pcjournal.471 507 \nPatrick, W.S., Link, J.S., 2015. Myths that Continue to Impede Progress in Ecosystem‐ Based 508 \nFisheries Management. Fisheries 40, 155–160. 509 \nhttps://doi.org/10.1080/03632415.2015.1024308 510 \nPearl, J., 2009. Causality. Cambridge University Press, Cambridge, UK. 511 \nPunt, A.E., Butterworth, D.S., de Moor, C.L., De Oliveira, J.A.A., Haddon, M., 2016. 512 \nManagement strategy evaluation: Best practices. Fish Fish. 17, 303–334. 513 \nhttps://doi.org/10.1111/faf.12104 514 \nPunt, A.E., Thomson, R., Little, L.R., Bessell-Browne, P., Burch, P., Bravington, M., 2024. 515 \nIncluding close-kin mark-recapture data in statistical catch-at-age stock assessments 516 \nand management strategies. Fish. Res. 276, 107057. 517 \nhttps://doi.org/10.1016/j.fishres.2024.107057 518 \nRubbens, P., Brodie, S., Cordier, T., Destro Barcellos, D., Devos, P., Fernandes-Salvador, J.A., 519 \nFincham, J.I., Gomes, A., Handegard, N.O., Howell, K., Jamet, C., Kartveit, K.H., 520 \nMoustahfid, H., Parcerisas, C., Politikos, D., Sauzède, R., Sokolova, M., Uusitalo, L., 521 \nVan den Bulcke, L., van Helmond, A.T.M., Watson, J.T., Welch, H., Beltran-Perez, O., 522 \nChaffron, S., Greenberg, D.S., Kühn, B., Kiko, R., Lo, M., Lopes, R.M., Möller, K.O., 523 \nMichaels, W., Pala, A., Romagnan, J.-B., Schuchert, P., Seydi, V., Villasante, S., Malde, 524 \nK., Irisson, J.-O., 2023. Machine learning in marine ecology: an overview of techniques 525 \nand applications. ICES J. Mar. Sci. 80, 1829–1853. 526 \nhttps://doi.org/10.1093/icesjms/fsad100 527 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n7 \n \nSaqib, M., Khokher, M.R., Yuan, X., Yan, B., Bearham, D., Devine, C., Untiedt, C., Cannard, T., 528 \nMaguire, K., Tuck, G.N., Little, L.R., Wang, D., 2024. Fishing event detection and 529 \nspecies classification using computer vision and artificial intelligence for electronic 530 \nmonitoring. Fish. Res. 280, 107141. https://doi.org/10.1016/j.fishres.2024.107141 531 \nSrivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014. Dropout: a 532 \nsimple way to prevent neural networks from overfitting. J Mach Learn Res 15, 1929–533 \n1958. 534 \nStawitz, C.C., Haltuch, M.A., Johnson, K.F., 2019. How does growth misspecification affect 535 \nmanagement advice derived from an integrated fisheries stock assessment model? Fish. 536 \nRes. 213, 12–21. https://doi.org/10.1016/j.fishres.2019.01.004 537 \nStock, B.C., Ward, E.J., Eguchi, T., Jannot, J.E., Thorson, J.T., Feist, B.E., Semmens, B.X., 538 \n2020. Comparing predictions of fisheries bycatch using multiple spatiotemporal species 539 \ndistribution model frameworks. Can. J. Fish. Aquat. Sci. 77, 146–163. 540 \nhttps://doi.org/10.1139/cjfas-2018-0281 541 \nSugeno, M., Munch, S.B., 2013. A semiparametric Bayesian approach to estimating maximum 542 \nreproductive rates at low population sizes. Ecol. Appl. 23, 699–709. 543 \nhttps://doi.org/10.1890/12-0453.1 544 \nThorson, J.T., 2019. Guidance for decisions using the Vector Autoregressive Spatio-Temporal 545 \n(VAST) package in stock, ecosystem, habitat and climate assessments. Fish. Res. 210, 546 \n143–161. https://doi.org/10.1016/j.fishres.2018.10.013 547 \nThorson, J.T., Anderson, S.C., Goddard, P., Rooper, C.N., 2025. tinyVAST: R Package With an 548 \nExpressive Interface to Specify Lagged and Simultaneous Effects in Multivariate Spatio-549 \nTemporal Models. Glob. Ecol. Biogeogr. 34, e70035. https://doi.org/10.1111/geb.70035 550 \nThorson, J.T., Ono, K., Munch, S.B., 2014. A Bayesian approach to identifying and 551 \ncompensating for model misspecification in population models. Ecology 95, 329–341. 552 \nhttps://doi.org/10.1890/13-0187.1 553 \nTriebe, O., Hewamalage, H., Pilyugina, P., Laptev, N., Bergmeir, C., Rajagopal, R., 2021. 554 \nNeuralProphet: Explainable Forecasting at Scale. 555 \nhttps://doi.org/10.48550/arXiv.2111.15397 556 \nWesselkamp, M., Moser, N., Kalweit, M., Boedecker, J., Dormann, C.F., 2024. Process-557 \nInformed Neural Networks: A Hybrid Modelling Approach to Improve Predictive 558 \nPerformance and Inference of Neural Networks in Ecology and Beyond. Ecol. Lett. 27, 559 \ne70012. https://doi.org/10.1111/ele.70012 560 \nWikle, C.K., Zammit-Mangion, A., 2023. Statistical Deep Learning for Spatial and 561 \nSpatiotemporal Data. Annu. Rev. Stat. Its Appl. 10, 247–270. 562 \nhttps://doi.org/10.1146/annurev-statistics-033021-112628 563 \nYang, H., Chen, C., Xue, X., 2025. Coupled convolutional neural network with long short-term 564 \nmemory network for predicting lake water temperature. J. Hydrol. 655, 132878. 565 \nhttps://doi.org/10.1016/j.jhydrol.2025.132878 566 \nZainol Mustafa, Nadia, 2025. Tweedie Distribution: A Statistical Solution for Unusually 567 \nDispersed Data. Sci. J. Stat. Probab. Its Appl. 3, 29–37. 568 \nhttps://doi.org/10.24127/sciencestatistics.v3i1.8003 569 \n 570 \n 571 \n  572 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n8 \n \nTables 573 \n 574 \nTerm in statistical ecology Term in neural networks Brief description \ndeterministic latent variable layer A hidden but non-random \nstate predicted from the \nmodel \nnonlinear transformation for \nelement of a latent variable \nneuron A method of abstraction to \nmap unobserved effects to \nresponses, for example, \nthrough logistic or \nexponential curves \nreverse-mode automatic \ndifferentiation \nbackpropagation Algorithm to efficiently \ncompute gradients by \nworking backwards through \nthe model \nparameter weight A model value controlling the \nstrength of an input's effect \nparameter estimation training The process of adjusting \nparameters or weights to \nminimize error on observed \ndata \nregression modelling supervised learning Predicting outcomes from \ninputs using observed data \ngradient-based optimizer2  Stochastic gradient descent \n(e.g., Adam) \nMethodology for updating \nmodel parameters during \ntraining to minimize error \noptimizing a management \npolicy \nstochastic policy ascent A process of adjusting policy \nparameters to optimize a \nspecific reward function \nnumerical overflow and \nunderflow \nExploding or vanishing \ngradient \nBecause the gradient \ncalculation involves products \nof terms, and each term in the \nproduct might be very large \nor small, the gradient might \n                                                 \n2 Where each optimization step is based on a subset of data; hyperparameters control how the \nexplored parameters update given the gradient, and each loop through all partitions of the data is \ncalled an “epoch” \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n9 \n \nend up being smaller or larger \nthan the numerical minimum \nor maximum allowed.  This \ncan preclude accurate \ncalculation of the gradient. \n 575 \nTable 1. Non-exhaustive correspondence of vocabulary from statistical ecology to neural 576 \nnetwork/AI methods. 577 \n 578 \n 579 \n Reinforcement Learning for \nFisheries \nManagement Strategy \nEvaluation \nRepresentation of \npopulation/system dynamics \nOperating model conditioned \nto historical data, termed \n“environment” \nOperating model conditioned \nto historical data \nData source used for model \nselection \nFuture simulation or historical \ndata  \nAlways future simulation \nRepresentation of future \nstochasticity \nSimulated variation in future \nrecruitment deviations \nSimulated variation in future \npopulation processes, \ntypically recruitment \ndeviations \nRepresentation of \nobservation and/or \nestimation uncertainty \nCan be addressed with \nspecialized neural network \narchitectures (i.e., \nenvironmental uncertainty as \nhidden state) \nHandled explicitly by \nestimation model \nHow catches are selected The policy function, which is \nlearned from direct \ninteractions with the \nenvironment, prescribes the \nannual catch according to the \nobserved state \nPre-defined harvest control \nrule sets annual catches based \non the current state (which \nmay be one or more estimated \nquantities or observed data) \nHow performance is \nmeasured \nreward function is maximized \nduring training as the agent \ninteracts with the \nenvironment; performance \nmetrics can be calculated after \ntraining based on the \noperating model time series, \nAfter simulations, calculate \nperformance metrics using \noperating model time series \nduring projection period \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n10 \n \nas in MSE \nComparison across policies Harvest policies compared \ninternally as part of learning. \nNovel harvest control rules \nexplored and evaluated; can \nconsider distinct yet flexible \nrules; no consideration of \nestimation models \nCan consider distinct yet pre-\ndefined harvest control rules \nand/or estimation models as \nstrategy components \nTable 2. Comparison of key concepts in reinforcement learning-based policy discovery and 580 \nManagement Strategy Evaluation. 581 \n 582 \n 583 \nFor use in projections and management modeling \n●  Does including neural-network derived projections of weight-at-age improve \nmanagement performance? Can these approaches be extended to other key population \ndynamics (i.e., numbers at age of recruitment?) \n●  Neural networks are not natively designed to handle nor represent uncertainty in their \npredictions, which is something highly valued by the fishery management community \nand often a requirement of reporting. How should we represent uncertainty in \nprojection models that use neural networks to project one or more components? \n●  If a neural network can predict data almost perfectly, how closely must that data \nrepresent the system for the predictions to be acceptable for management?  \n●  Is it possible to construct a neural network that predicts stock biomass using a feature \narray of input data more accurately than a process-based estimation model?  \n●  If so, can such a model be used to explore the impacts of varying catch levels on future \npopulation biomass (and thus fully replace the process-based projection framework \nused to calculate management quantities)? \n●  Can neural networks support the simulation component of management strategy \nevaluation, perhaps enhancing the accuracy of predicted population dynamics, \nobserved data, or both?  \nFor use in quota setting \n●  Can the data predicted by a neural network replace true observational data for short-\nterm management use (for example, to inform an empirical harvest control rule in the \nabsence of a survey), and if so, how long is “short term”?  \n●  How might a management system integrate processed-based models with RL-derived \npolicies? At what frequency would operating models/environments need to be updated \nto confidently rely upon a policy learned by RL? \n●  How would the development of RL-based policies be communicated to the fishery \ncommunity, and how would stakeholder engagement change under this paradigm? \nFor the fisheries science community \n●  Are there historical, social or economic characteristics of a fishery and management \nsystem that are not well suited for these applications?  \n●  What are the ethical considerations of providing management advice based on AI \nmodels that inherently offer less explainability than existing methods?  \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n11 \n \nTable 3:  List of questions arising from case studies, that could form the basis of future research 584 \nin scientific fisheries management. 585 \n 586 \n 587 \n 588 \n  589 \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n12\n \nFigures  590 \n 591 \n 592 \n593 \nFigure 1. Conceptual diagram of case studies. The case studies were selected and are ordered 594 \n(from left to right) in terms of how tractably each NN application could be included in current 595 \nscientific fishery management frameworks.  596 \n 597 \n 598 \nFigure 2. Case Study 1: Number of times each process- and neural network based model resulted 599 \nin the lowest average root mean squared error (RMSE) when fit to simulated data (n = 300) 600 \nwithout (top row) or with (bottom row) a temporal trend in true fish growth. Hindcast represents 601 \n10-fold leave year out cross-validation and “y+1” and “y+2” represents forecast skill from 602 \nsequential peels of historical data and forecasted for two future years. 603 \n12 \n \ned \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n13\n \n604 \n Figure 3. Case Study 2: Investigation of neural networks for spatio-temporal survey data 605 \ninterpolation and index standardization. A) Mean scaled annual indices (lines) and 95% 606 \nconfidence interval (shaded ribbon) for 25 replicates of a simulated 12-year biomass time series; 607 \nblack solid line is true biomass whereas all other colors are model estimates. B) Performance 608 \nmetrics of various estimation models (colors) in terms of RMSE in biomass (lefthand figure) or 609 \ndensity (righthand side) for estimates in a 20x20 gridded domain for 12 simulated years across 610 \n25 replicates. C) Maps of true and estimated log biomass across the domain for a single replicate 611 \nand year; results shown only for models that produce an interpolated surface as part of the 612 \nestimation step. 613 \n  614 \n13 \n \ns; \nte \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n14\n \n 615 \n Figure 4. Case Study 3: Investigation of neural-network based reinforcement learning method 616 \n(orange lines and points) for fisheries management, in comparison to 25 replicates of an MSE 617 \n(blue lines and points) for EBS pollock using an OM conditioned on observations from 1964-618 \n2024 (grey rectangle; years are compressed) with a projection period from 2025-2050. Top row: 619 \nrealized catch (millions of mt) under the constant (1a) and Ricker (1b) recruitment scenarios. 620 \nMiddle row: stock spawning biomass in the operating model under the constant (2a) and Ricker 621 \n(2b) recruitment scenarios. Horizontal dashed line indicates 20% of unfished biomass and the 622 \ncolored line represents the median across simulation replicates for a given year. Colored ribbons 623 \nrepresent 95% simulation interval. Bottom row: realized harvest policy in terms of the quota 624 \n(millions of mt) specified for future years(s) (points) versus the observed Bottom Trawl Survey 625 \nBiomass (t) arising from each method under the constant (3a) or Ricker (3b) recruitment 626 \nscenarios. The MSE did not use a survey-based method for setting quotas, but observations from 627 \n25 MSE replicates and the subsequent years’ catch are shown for comparison. 628 \n 629 \n14 \nm \n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint \n\n105 and is also made available for use under a CC0 license. \n(which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC \nThe copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint","source_license":"Public-Domain","license_restricted":false}