二次奖罚学习自动机

Quadratic Reward-Penalty Learning Automaton

研究了奖罚型学习自动机的一种非线性强化算法。与线性的奖罚模型（ＬＲＰ）不同，新模型的行动选择概率的更新函数为二次的。这使得该模型的学习性能优于ＬＲＰ，且对不同的环境，其具有不同的行为和特点。

In this paper a nonlinear reinforcement algorithm for reward penalty type learning automata is studied. It is different from the linear reward penalty model (L RP ), the update function of action selection probability of the presented algorithm is quadratic. The learning performance of the new model is superior to the one of the L RP Additionally, for different environments, the proposed automaton possesses different behaviours and properties.