Solving Maze Problem with Abstraction Selection

Abstraction has been extensively studied in the fields of artificial intelligence. It is especially useful for highdimensional continuous domains, since abstraction can reduce the large state space and action space in such problems. In this project, we solve the multi-task Maze problem by temporal difference learning under the abstraction selection framework. Simulation results show that the multi-task maze problem can be solved efficiently with the abstraction selection framework. We further compare the algorithm proposed in an existing literature with our modified algorithm. Results showing that the multitask maze problem can be solved more efficiently by using tabular form TD learning, compared with the existing function approximation framework proposed in [1].