Action chunking as conditional policy compression

Many skills in our everyday lives are learned by sequencing actions towards a desired goal. The action sequence can become a ``chunk'' when individual actions are grouped together and executed as one unit, making them more efficient to store and execute. While chunking has been studied extensively across various domains, a puzzle remains as to why and under what conditions action chunking occurs. To tackle these questions, we develop a model of conditional policy compression—the reduction in cognitive cost by conditioning on an additional source of information—to explain the origin of chunking. We argue that chunking is a result of optimizing the trade-off between reward and conditional policy complexity. Chunking compresses policies when there is temporal structure in the environment that can be leveraged for action selection, reducing the amount of memory necessary to encode the policy. We experimentally confirm our model's predictions, showing that chunking reduces conditional policy complexity and reaction times. Chunking also increases with working memory load, consistent with the hypothesis that the degree of policy compression scales with the scarcity of cognitive resources. Finally, chunking also reduces overall working memory load, freeing cognitive resources for the benefit of other, not-chunked information.