Abstract
In real world applications, learning algorithms often have to handle several dozens of actions, which have some distance metrics. Epsilon-greedy or Boltzmann distribution exploration strategies, which have been applied for Q-learning or SARSA, are very popular, simple and effective in the problems that have a few actions, however, the efficiency would decrease when the number of actions is increased. We propose a policy function representation that consists of a stochastic binary decision tree, and we apply it to an actor-critic algorithm for the problems that have enormous similar actions. Simulation results show the increase of the actions does not affect learning curves of the proposed method at all.
