Abstract: To improve the classic reinforcement learning, through the introduction of motivation, prior knowledge is introduced, and the learning speed is speeded up. As to the iteration strategy, it adopts “on-policy” iterative Sarsa learning algorithm instead of traditional “off-policy” Q learning algorithm. It proposes Multi-Motivation Sarsa learning algorithm（MMSarsa） and respectively carries out the comparative tests on tank battle simulation with Q-learning algorithm and Sarsa learning algorithm. The results of experiment show that Sarsa learning algorithm based on motivation guidance has fast convergence rate and high learning efficiency.

Key words: multi-motivation guidance, Q learning, Sarsa learning, unmanned tank, battle simulation

摘要： 对标准的强化学习进行改进，通过引入动机层，来引入先验知识，加快学习速度。策略迭代选择上，通过采用“同策略”迭代的Sarsa学习算法，代替传统的“异策略”Q学习算法。提出了基于多动机引导的Sarsa学习（MMSarsa）算法，分别和Q学习算法、Sarsa学习算法在坦克对战仿真问题上进行了三种算法的对比实验。实验结果表明，基于多动机引导的Sarsa学习算法收敛速度快且学习效率高。

关键词: 多动机引导, Q学习, Sarsa学习, 无人坦克, 对战仿真