This paper describes a reinforcement learning algorithm based on super
vised learning. Many of reinforcement algorithms use associative searc
h to discover and learn actions that make the system perform a desired
task. One problem with associative search is that the system's action
s are often inconsistent. In the searching process, the system's actio
ns are always decided stochastically. Therefore, the system cannot per
form learned actions more than once. To solve this problem, this algor
ithm uses a neural network which can predict an evaluation of an actio
n and control the influence of the stochastic element. The effectivene
ss of this algorithm was checked by computer simulations.