Quagent control via Passive and Active Learning

Artificial intelligence algorithms using passive and active learning versions of direct utility estimation, adaptive dynamic programming and temporal difference approaches to simulate an agent. The explored worlds consisted of discrete states (positions) bounded by internally generated “walls” that included one or more terminal states and a pre determined configuration of rewards for each state. Passive learning algorithms use a pre-calculated optimal movement policy to travel from the start state to the best terminal state. Active learning algorithms use an initially random movement policy and correct the known policy based on the percepts received within the world. In a passive learning scenario, all three approaches were found to be effective since the optimal policy was known. In active learning scenarios, direct utility estimation tends to result in unsolvable situations or less than optimal policies as a result of poor initial random that create unreachable areas. Adaptive dynamic programming is very effective in an active learning scenario but is inefficient in storage space and often fails to evaluate the entire map. Temporal difference learning approaches are very space efficient but often require more trial runs to approach the same level of accuracy that adaptive dynamic programming achieves. These different learning techniques can be tested with or without the use of quake and the Quagent bot.