QTRB: TEAM-BASED REGION BUILDING USING Q-LEARNING TO DERIVE POLICY ON PROGRAMS PARAMETERIZED BY LOCAL REWARD SIGNAL

Abstract

While attempting to solve 2-dimensional grid world maze tasks, it was observed that genetic programming is limited by its random initialization and no use of local reward. This thesis proposes a hybrid algorithm called QTRB, team-based region building with q-learning, which attempts to integrate genetic programming and reinforcement learning to use local reward during evolution. During evolution, QTRB constructs programs based directly on local environmental reward; programs are then passed to a reinforcement learning agent to learn on as a model. QTRB was tested to solve variously sized 2-dimensional maze tasks, hypothesizing that policy can be derived from an agent learning from this model. The results suggest that QTRB can derive policy on the given tasks, with fewer direct environment queries than traditional q-learning as the task size scales.