Non-symmetric Preferences in the IPA Market with Reinforcement Learning

Machine Learning has been proposed to support and optimize market-based resource allocation. In particular, reinforcement learning (RL) has been used to improve the allocation in terms of the utility received by resource requesting agents in the iterative price adjustment (IPA) mechanism. In such an approach, utility functions describe the agents' preferences for resource attributes and are the basis for RL to learn demand functions that are optimized for the market. It has been shown that the reward functions based on the individual utility of the agents and the social welfare of the allocation can deliver similar social results when the market consists only of learning agents with symmetric preferences. In this paper we investigate the IPA market-based resource allocation with RL for the case of agents with non-symmetric preferences. We show through experimental investigation that the results observed above are also held in this case. In particular, we show that the individual-based reward function is able to approximate the solution to the fairest Pareto-optimal allocation in situations where the social-based reward function fails.