一种基于新奇的动作发育模型

An Action Developmental Model Based on Novelty

机器人的动作是一切活动的基本单元。就足球机器人而言,好的动作设计实现是决策实现的重要保证。传统的强化学习模型在整个学习过程中使用恒定学习速率,导致在未知环境下收敛速度慢,且适应性差。针对以上问题,提出了一种新的动作发育模型——基于新奇的动作发育模型;该模型在学习过程中使用基于状态的遗忘均值的学习速率,更加符合人类发育的真实过程。模型采用内在价值系统,该系统由三部分组成:奖励、惩罚和新奇评判。在机器人足球比赛中,通过机器人截球实验表明,该模型在不断变化的环境下可以高效而准确地完成相应的截球动作。

The robot＇s action is the basic element of the activities,for the robot,good action design is the important pledge to implement strategy.The learning process uses the constant learning rate in the traditional reinforce learning model,because of that robot learn in a low convergence speed and with the poor adaptation.For the above questions,a new kind of an action developmental model-action is proposed developmental model based on novelty.The model in the learning process uses the learning rate which based on the amnesic average,which is consistent with human real development process.This model uses innate value system which is consists of three parts： reward,punishment and the novelty.Robots intercepting experiments indicates that the model can be efficiently and accurately to carry out appropriate actions in constantly changing environment.