Research On Game Intelligence Based On Behavior Tree

Finite state machine is rigid and not flexible,while fuzzy state machine brings reasonable randomness and lacks adaptability.Behavior trees gradually have become the main trend of game artificial intelligence.'However traditional behavior trees are very difficult to mix machine learning techniques together.The strategy decision?tactics decision and behavior level decision only be expressed in a sole behavior tree while team artificial intelligence designed,which will result in low efficiency of decision making and weak software robustness.This thesis studies how to learn online and discusses the team intelligence in the behavior tree about which we present hierarchical behavior tree architecture,the main work is as follows:A kind of method which combines pretreatment and online learning based on behavior tree is put forward.Q-Condition nodes are introduced to replace the original condition node of behavior trees and update selection weights of nodes with Q-learning for online learning,which injects reasonable randomness into behavior decision.A hierarchical decision-making architecture based on behavior trees is proposed,through which we can design appropriate behavior trees according to characteristics of different layers,and between layers messages or commands can be passed to do efficient communication which mechanism is also the foundation of the event or message driven behavior tree.Under this architecture,behavior trees can automatically run and be interrupted by orders or messages.After that,corresponding processing program will be performed immediately.At the same time,hierarchical behavior trees avoid redundant single tree decision and improve the robustness of the system.This paper implements a communication interface of interactive platform based on Unreal Development Kit and completes a behavior tree system based on UDK.Lastly,Team decision-making artificial intelligence system using three layers architecture is implemented.