Applying SAQ-Learning Algorithm for Trading Agents in Bilateral Bargaining

Abstract

In this research we use a learning method called SAQ-Learning to use for agents in a single-issue bargaining process. SAQ-Learning algorithm is an improved version of Q-Learning algorithm that benefits from the Metropolis criterion of Simulated Annealing (SA) algorithm to overcome the challenge of finding a balance between exploration and exploitation. Q-Learning is one the most important types of Reinforcement Learning (RL) because of the fact that it does not need the transition model of the environment. Artificial Intelligence (AI) approaches have attracted interest in solving bargaining problem. This is because Game Theory (GT) needs some unrealistic assumptions to solve bargaining problem. Presence of perfectly rational agents is an example of these assumptions. Therefore by designing SAQ-Learning agents to bargain with each other over price, we gained higher performance in case of settlement rate, average payoff, and the time an agent needs to find his optimal policy. This learning method can be a suitable learning algorithm for automated online bargaining agents in e-commerce.