One of the limitation of IPD is that there are only really two strategies that the actors can take. Obviously, in real life there are a lot more and this paper I read today begins to discuss this. They had a large network of actors and played the Ultimatum Game (which I will talk about some day) many times with neighbors in their network. Furthermore, they were able to adjust their strategies so as to converge and maximize their payoff, and thus learn how to deal with other individuals. The continuous nature of the strategy space is found in the fact that they can offer or accept offers at not just discrete levels, such as in IPD (defect or cooperate) but over a range of values in between the minimum and maximum offers allowed.
The authors also investigate how to model fair decision which humans seem to do quite quickly. He also denigrated the impact of reputation a little but I think that that still comes into play. Humans are much more emotional than computer agents and so I'll have to find another paper discussing that issue in more depth.
Reference:
"Learning to Reach Agreement in a Continuous Ultimatum Game"
de Jong, Steven et al.
JOURNAL OF ARTIFICIAL INTELLIGENCE RESEARCH Volume: 33 Issue: Pages: 551 Published: 2008
Subscribe to:
Post Comments (Atom)
No comments:
Post a Comment