An Optimal Transfer of Knowledge in Reinforcement Learning through Greedy Approach

Transfer Learning has been used in Reinforcement Learning methods by integrating information from preceding task. Although, these algorithms generally have to be assigned either a complete prototype of the task or direct mapping of information from input task to output task. The agent may not have the knowledge of such complex information. However, it is able to examine the resemblance between two tasks which are similar in taking action or approach. In this paper efforts has been made to solve mountain car problem (a famous problem domain in the area of Reinforcement Learning) using the SARSA ( $\boldsymbol{\lambda}$ ) [10] algorithm and MASTER [3] method by using greedy approach. The greedy approach has been used in applying parameters for transferring the knowledge within 2-dimensional mountain car to 3-dimensional mountain car. In the existing techniques the best results are obtained with use of parameters like epsilon decay factor, eligibility traces with greater decay factor and positive weight initialization. In the article, greedy approach has been applied with these parameters to get optimal results.