Implementing Realistic Human Motion in Games

The computer games industry is increasingly being pressured to create games that are realistic, in the sense that rules that apply in the real physical world also do apply in computer games. Human motion (i.e. running and walking) is one area that is prevalent in computer games especially in genres such sports, action, and adventure. In real life, motion is a key element for all living beings if they are to engage. It is primarily governed by thought and actuated by muscles and their energy sources. Complex energy sources in muscles are used to produce force, which in turn creates motion. Currently, most games ignore, cheat, or implement rule-based logic to reflect the thought and muscle processes. In these cases, motion is not realistic. This thesis attempts to model realistic human motion governed by realistic energy constraints, but can which realistically achieve long-term goals in a competitive real world environment. An athletic event is used to construct a realistic competitive environment that allows us to focus on the energy constraints and capture the motion dynamics. Q-learning, a reinforcement learning method in machine learning, is then be used to model domain knowledge acquisition necessary for the decision making processes that produce realistic behavior for motion. Generalization of the behavior for larger environments using a combined Q-learning and neural networks model is then discussed implemented and measured. Based on the theoretical analysis and the empirical results from the construction of the athletic event, this thesis concludes that a realistic model of the energy dynamics that control the way we execute motion be realized and implemented whilst also allowing them to generalize due to evidence the learning shows convergence. This model can then be used in computer games to increase the realism in chase and evasion strategies.