Real-time ai in xpilot using reinforcement learning
Martin W. Allen, Kristen Dirmaier, Gary B. Parker · World Automation Congress · 2010
Reinforcement learning (RL) allows agents to learn a best-possible long-term course of action, based on immediate positive and negative rewards. This approach enables real-time learning, since the agent constantly adjusts the value of actions taken, eventually selecting that action with highest value in each environment-state it encounters. We investigate the use of the Q-learning RL technique in an agent that learns to intelligently navigate the Xpilot video game environment in real time. We compare learning performance for different reward and action models, and discuss the challenges of RL methods in such a reasonably complex domain.