Improved Trial and Error Learning and Application to Distributed Dynamic Channel Assignment

This paper addresses the distributed dynamic channel assignment (DDCA) problem in wireless clustered networks using a fully distributed multi agent reinforcement learning (MARL) approach named trial and error learning (TEL). By fully distributed, we mean that there is no communication between the clusters and as a consequence they have only access to local information to chose their channel. The TEL algorithm has the interesting property to converge to a solution with global optimality guarantee in such context, and has already been used to solve the DDCA problem. In this paper, we propose three heuristics to improve the convergence speed and the stability of the TEL algorithm, resulting into the improved TEL (ITEL) algorithm. We are able to prove that the good convergence property of the TEL algorithm still holds for the ITEL one. We provide simulation results that show the performance gains of the ITEL algorithm as compared with the conventional TEL. We also perform an ablation study to identify which heuristics are the most important among the proposed ones.