Multi-agent-Based Systems in Machine Learning and Its Practical Case Studies

Abstract

In the area of Decision Science research, Multi-Agent Systems (MAS) has been hailed as a new paradigm for conceiving and designing software systems. The goal of this chapter is to underscore the need for and usefulness of MAS by giving the reader an insight into the agents’ characteristics, its interaction with its environments, various performance measures and different types of MAS. This chapter gives insights on Reinforcement Learning Agents the root of which originates from animal learning theory. In reinforcement learning, the agents comprehend the state of the environment at each step, and adopt the decision on the action to be taken, causing its environment to transit into a new state. The agents each receive a scalar reward that evaluates if the transition made is aligned with the goal. The novelty of this chapter is the use of a multi-agent programmable tool called NetLogo to explain the reinforcement learning technique with appropriate examples and the results obtained are thoroughly analyzed. Prominent algorithms like Policy Hill-Climbing, Q-learning, and its variants such as Nash Q, Minimax Q-Learning, and optimization techniques such as Distributed Constraint Optimization (DCOP) in multi-agent systems which help the agents in achieving equilibrium are discussed. The chapter also focuses on the typical domains that can benefit from MAS. It comprises of instances from examples of MAS implemented using the NetLogo models for enhancing the understanding of agents and their environment. The chapter concludes by giving the reader a sneak-peek into the latest research trends in MAS.