Distributed Cooperative Learning Over Networks via Fuzzy Logic Systems: Performance Analysis and Comparison

This paper studies a distributed machine learning problem by applying a distributed optimization algorithm over an undirected and connected communication network. Each node has its own fuzzy logic system (FLS) based machine whose weights are trained by the proposed FLS-based distributed cooperative learning (DCL) algorithm to reach the optimum of the global cost function. The training process utilizes the data that are distributed among different nodes and cannot be gathered at any node in the network. The main advantages of the FLS-based DCL algorithm are as follows: It has an exponential convergence; it requires a small amount of computation and communication at each iteration step; and the private and confidential information is protected without exchanging raw data between neighboring nodes. These advantages are verified by performing simulation experiments to compare the FLS-based DCL algorithm with the distributed average consensus based learning algorithm, the alternating direction method of multipliers based learning algorithm and the diffusion least-mean square algorithms.

I. Introduction

This paper investigates a distributed machine learning problem in an undirected and connected communication network based on fuzzy logic systems (FLSs). This problem is a type of distributed in-network data processing problem, where all nodes in the network cooperatively find an identical but unknown pattern only by sharing learned knowledge with their neighboring nodes.