A Pruning technique to increase the rate of convergence of Learning Algorithms

摘要

The rate of convergence of a learning automaton depends on the number of actions. When the number of actions are large the automaton becomes slow due to many updates to be made at each stage. In this paper we suggest a method to increase the rate of convergence by pruning the action set. We mimic the human tendency of pruning the choices available as more information becomes available i.e., dropping actions with small action probabilities from further consideration. The learning algorithm considered is a P-type reward-penalty learning algorithm. A truncation rule for dropping an action is also given. Simulation studies were conducted to show that a learning algorithm with pruning technique converges faster than a similar algorithm without pruning. The motivation, benefits and issues related to pruning are discussed. The idea of pruning the action set can be extended to all types of learning algorithms to increase the rate convergence.