Search Papers | Poster Sessions | All Posters
Poster C108 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Optimal Learning in Temporally Structured Environments
Niloufar Razmi1 (), Matthew Nassar1; 1Brown University
Biological agents live in a dynamic world and can exploit structure in their environment to improve the efficiency of learning. Previous work has yielded normative learn- ing algorithms that prescribe learning strategies for spe- cific environmental structures, but leave open the ques- tion of how humans and animals might infer the structure of their current environment. In this project, we propose an optimal theoretical model of learning the structure of varying environments. Specifically, we define learning the structure of change as putting a prior on the transition matrix of a hidden Markov model and using observations to update that prior with Bayes rule. With minimum as- sumptions imposed on the generative model of the en- vironment statistics, we test our model in four different environments and find signatures of context-appropriate behavior previously observed in humans. Our work proposes the first unifying model of adaptive learning through experience in complex temporally structured en- vironments.
Keywords: State representation Structure learning Bayesian inference Hierarchical Dirichlet process