Abstract:
An intelligent agent should possess the capability of solving problems related to the task of interest based on the perception of its virtual environment acquired from its past and present interactions. These agents should be able to extract the fundamental trait of being intelligent in order to possess human behavior. Learning and planning are the major modalities that contribute to this trait. Brown-UMBC Reinforcement Learning and Planning (BURLAP) is an existing library that comprises of algorithms that help the agent imitate the planning and learning behaviors of a human being. The algorithms in BURLAP can be used to implement intelligent agents in virtual worlds including Minecraft as it offers challenges that of a real-life platform. Minecraft allows the use of mods which are modifications to the environment based on the user’s preference. The mod, BurlapCraft can be used to deploy the algorithms present in BURLAP. It includes scenarios such as dungeons that are of different caliber to test these algorithms. In literature, the developers of BurlapCraft have tested Rmax, Breadth First Search (BFS) and A star (A*) but have not implemented algorithms, Iterative Deepening A star (IDA star), Depth First Search (DFS), Q learning and State Action Reward State Action (SARSA) in BURLAP which makes the potential benefits of these algorithms unknown.
This research focuses on testing the efficiency and effectiveness of the reinforcement learning and planning algorithms, Q learning, SARSA, IDA star and DFS developed in BURLAP using the mod, BurlapCraft to make certain of their potential in solving a task oriented problem. It further analyses the potential of applying these algorithms in a pre-designed scenarios that are of different caliber which in turn would lead to the selection of the best fit and worse fit algorithms for the respective problems.
The performance evaluation identified that IDA star and Q learning algorithms do make an impact in improving the efficiency of the agent in completing the specified task. It also identified the best fit and the worst fit algorithms for the respective scenarios that could be mapped to general Artificial Intelligence (AI) related problems such as decision making, traversal and search present in the real world.
