Sample-Efficient Behavior Cloning using Hierarchical Event Memory of Bayesian Networks

NeurIPS 2024 Workshop BDU Submission44 Authors

Keywords: behavior cloning, Bayesian network, event memory, temporal model, sample

TL;DR: This paper introduces a behavior cloning agent that uses event memory to rectify class imbalances in the training dataset and efficiently approximate the expert policy.

Abstract: When humans imitate others they often rely on their memory of individuals demonstrating the desired behaviors to emulate. This not only permits people to reproduce taught behavior, but also enables them to generalize procedural expertise to novel situations. This paper introduces HEMS-BC; a behavior cloning agent that uses a novel, psychologically plausible model of event memory. From sequential observations of an expert, the memory system represents states, observations, and state transitions as Bayesian causal models, and stores them into a hierarchically organized event memory taxonomy. In response to observation queries, posterior conditional samples of observation-action pairs are drawn to rectify class imbalances in the training dataset that the imitation learner uses to approximate the expert policy. Our findings show that our method achieves and maintains expert-level performance from fewer expert demonstrations compared to a baseline system with no event memory capabilities.

Submission Number: 44