Agent YK: An Efficient Estimation of Opponent's Intention with Stepped Limited Concessions

摘要

On my agent, I heavily focused on this efficiency issue to realize a great scalability for the estimation of opponents' preferences. Naturally, to realize it, we should consider some approximation approaches. In my agent, I rather utilized a local-search based approach for searching good bid alternatives with a limited approximate estimation of the opponent's utility space. This heuristic offering approach could also be effective to produce a better outcome as well as a better social welfare when the opponent has a mechanism which tries to sense the degree of intention for a cooperation. However, due to its nature of approximation, it may not always be sensed as a cooperative even when the agent has been tried to be cooperative to the opponent. Especially, on the nearly ending phase of the negotiation, both the agent and the opponent may stick their best compromised proposals together due to this issue. To overcome such a situation, I introduced a special gimmick named "Hotoke no kao mo sandomade", that allows more compromised proposals at most three times. This gimmick was named that the agent will "forgive" the uncooperative reactions of the opponent when the opponent stops such uncooperative reactions after a compromised proposal from the agent was received. The approach is based on my own analysis of the possible behaviors of the agents to be submitted as well as the domains, and of course respect the same proverb that is popular in Japan.