Cette page appartient aux archives web de l'EPFL et n'est plus tenue à jour.
This page belongs to EPFL's web archive and is no longer updated.

Interacting with Humans II

You are strongly encouraged to read the following paper.

 

J. Pita, M. Jain, M. Tambe, F. Ordonez, and S. Kraus, 2010, Robust solutions to Stackelberg games: Addressing bounded rationality and limited observations in human cognition, Artificial Intelligence, Vol 174, pp. 1142-1171.

Posted by Katherine Skirving Larson on Thursday 5 April 2012 at 9:47
Comments
The paper is a journal article, an extension of their previous publications in 2009.

In this article, they incorporate the nature of human: bounded rationality (deviate from the expected optimal strategy)and limited observation (about leader's strategy, giving them a false impression of that strategy) into account when solving the Stackelberg game, i.e. obtain the highest payoff for the leader. This is the main contribution. The key finding is that when we do not consider these humanly characteristics of the followers (the adversaries), the leader will obtain such an unexpected (could be much worse) outcome. Despite providing the theoretical foundation, the authors also supply a nice experiment result and analysis.

Key observation:
- The \epsilon given as a parameter in their approach, COBRA(\alpha, \epsilon) is the fraction of optimal response that the player will choose. Which means that when we define \epsilon we have already make an assumption on the end result on the limited observation (without modeling the limited observation itself). Isn't that strange? First when I read about limited observation, I am thinking about Level-k as in "Beyond..." paper by Leyton-Brown.
- As in the two papers appear in the blog discussion earlier "Beyond..." and "Behavioral..." by Leyton-Brown, these approach suffer from parameter setting problem.
- In experiment determining \alpha for unlimited observation, their key point about strategy entropy is insightful.

Some concern:
- The initiative about conducting the thorough experiment is good enough, but I'm afraid that we also have an effect of HCI in the display of triplet-list. Since for unsophisticated person, observing triple-integer will be difficult compare to if we present the information in rows of images 8-doors with or without guards.
- Unfortunately the background of the human subject is not describe here. As mentioned in previous point, for unsophisticated user observing integer will be hard to perform -> images is easier.

Future work:
- As the authors suggest, determining the best setting for \alpha is the next step. This is really important since in the security domain, the cost of false negative is very high. So, still, building a high performance model (modeling adversary behavior and maximize leader payoff) is undoubtedly desirable.
Posted by Tri Kurniawan Wijaya on Tuesday 17 April 2012 at 14:48