Improving Sparse Reward Classifiers for Model-Based Reinforcement Learning in Low-Data Regimes

Uploaded version with minor typographic errors corrected, per request of the Grad office, 12/13/2022 (via email).Model-Based Reinforcement Learning consists of learning a model of an environment in which we can train an agent. This approach is useful when we are limited in the number of interactions the agent can take with the environment: learning a model enables to simulate additional interactions. However, when dealing with sparse-reward environments, teaching the model when to deliver rewards can be challenging: the model is built upon few interactions, of which only a small subset contains rewards. In this work, I show that different training methods involving the decoupling of the model’s reward classifier from its transition estimator considerably improves the performance and robustness in sparse-reward, low-data regimes. Additionally, I propose modifications to SimPLe [1] a model-based algorithm, and show that they increase its performance for this setup. Finally, I further improve these results by leveraging the model-based setup, in particular I introduce Model-Based Aware Reward Classification (MBARC), a training method that weighs training examples based on how close they are temporally from a reward.M.S