Learning from Reinforcements



We look at the task of learning from reinforcements. One-step connectionist (neural network) Q-Learning was implemented to act as a controller for the agent class GeislerPlayer. An instance of this player can be viewed in the AgentWorld program and is labeled as net enhance 2. The backpropogation algorithm is used as the policy function.

Summary of Results

To run the agent world, download the Windows 9x/NTAgentWorld.exe executable and the jar file, or download the following class files and run using jview or your favorite java application executor. For details behind the algorithm or suggestions email bgeisler@cs.wisc.edu. Please stay tuned, an applet version comming soon!