基于智能决策算法的社会演变游戏开发

It’s suitable to research issues of individual decision in primitive tribe with less factors. Our experiments manage to make transitions with behavior trees and get values of each designed action through a deep q-learning network trained by a reward function of reinforcement learning. Then relink all action modules according to the values to the root node of behavior tree. After more experiments, we provide several improvements such as prioritized experience replay and dynamic learning rate for the final expected human-normalized action decision.