Fuzzy Q-Learning for first person shooters

Here machine learning techniques in the context of their application within computer games is examined. The scope of the study was that of reinforcement learning [RL] algorithms as applied to the control of enemies in a video game dynamic environment thus providing interesting new experiences for different game players. The project proved reinforcement learning algorithms are suitable and useful for simulating intelligence of agents within a game. The implemented learning bot exhibited interesting capabilities to adapt to a human player playing style in addition to outperforming other implemented and downloaded bots. Test results showed that incorporating learning into a game can reduce the extensive scripting and tuning phase, while retaining the ability to guide the NPCs not to exhibit totally unrealistic or unexpected behaviors. This gives the designers the scope to explore new strategies. The effect of the exploration-exploitation policy on the learning convergence was studied and tested in depth. The combination of the two most popular reinforcement algorithms [Q-Learning and TD(λ)] resulted in faster learning rates and realistic behavior through the exploration period.