Enhancing Multi-agent Bargaining with the TD-based Reinforcement Learning Approach

This study proposes a negotiation mechanism that applies TD-based reinforcement learning to deal with on-line bargaining between two parties both with incomplete information. The agent embedded with the TD-based reinforcement learning capability can learn dynamic strategy incrementally by itself with the past bargaining experiences. This study investigates the scenarios that a TD-based seller agent bargains with buyers who have different risk-attitudes, and both seller and buyer agents gifted with the TD-based reinforcement learning capability to negotiate with each other. Bargaining experiments are conducted on JADE, a software agent framework based on the FIPA specifications, to evaluate the bargaining performance in average payoff and settlement rate. Results show that the negotiation mechanism handles the multi-agent bargaining process effectively. This study can be further applied to electronic commerce environment for on-line automated bargaining.