Reading Literature

High-level Reinforcement Learning in Strategy Games

To take a single agent approach which learns a best response strategy against a fixed player.

Both anecdotal and experimental evidence showing that humans often play fixed or slowly changing strategies. Even in simple games, it has been shown that humans often consider simplistic models of their opponents such as that they are playing against a fixed policy

Reinforcement Learning

Q-learning

Model based Q-learning

Factored state representation

These policies could win more often than both random policies and the hand-tuned (fixed) policy of the game AI.

Game: Civilization 4