Program synthesis approaches to improving generalization in reinforcement learning

Abstract

To perform real-world tasks, robots need to rapidly explore and model their environment. However, existing methods either explore slowly, are data-inefficient, or need to leverage significant prior knowledge that limits their generalization. In this thesis, we explore applications of techniques from the program synthesis literature to improve the generalization and data-efficiency of reinforcement learning agents. Two complementary approaches are explored. First, we explore leveraging program synthesis techniques to meta-learn exploration strategies, and automatically synthesize new explorations strategies competitive with state of the art benchmarks. Second, we explore applying program synthesis to the problem of learning factored world models and achieve promising preliminary results. We see these results as promising examples of the potential of integrating program synthesis techniques with the rest of our modern modern reinforcement learning and robotics toolkits, increasing generalization in the process.