Learning in automated negotiation

Renting, B.M.

doi:doi:10.60602/1887/4284788

This dissertation advances automated negotiation by developing agents that can learn and adapt across diverse negotiation settings through three increasingly sophisticated approaches: automated algorithm configuration, portfolio-based strategy selection, and end-to-end reinforcement learning with graph neural networks. We demonstrate that while each of our developed methods successfully pushes the boundaries of what negotiation agents can achieve, our end-to-end reinforcement learning approach shows particular promise in reducing human-induced bias while maintaining conceptual simplicity. In the second part, we evaluate negotiating agents through extensive empirical research, including organising the Automated Negotiating Agents Competition (ANAC) and demonstrate that learning agents generally outperform non-learning agents. Our analysis reveals fundamental limitations in standard evaluation metrics for negotiation agents, particularly showing that rankings based on average...Show moreThis dissertation advances automated negotiation by developing agents that can learn and adapt across diverse negotiation settings through three increasingly sophisticated approaches: automated algorithm configuration, portfolio-based strategy selection, and end-to-end reinforcement learning with graph neural networks. We demonstrate that while each of our developed methods successfully pushes the boundaries of what negotiation agents can achieve, our end-to-end reinforcement learning approach shows particular promise in reducing human-induced bias while maintaining conceptual simplicity. In the second part, we evaluate negotiating agents through extensive empirical research, including organising the Automated Negotiating Agents Competition (ANAC) and demonstrate that learning agents generally outperform non-learning agents. Our analysis reveals fundamental limitations in standard evaluation metrics for negotiation agents, particularly showing that rankings based on average utility are highly dependent on opponent group composition. The dissertation concludes by proposing multi-agent meeting scheduling as a concrete application domain that could provide clear performance criteria and drive meaningful progress in automated negotiation research.
Show less

Leiden University Scholarly Publications

View statistics

Documents

In Collections

Learning in automated negotiation

Funding