Learning from repeated games /
This dissertation makes two primary contributions to the literature. The first contribution is to report on experiments that examine incentive problems in dominance solvable games. A novel matching protocol, T-Death, is presented. This protocol was used with two dominance solvable games. Game I...
| Main Author: | |
|---|---|
| Format: | Thesis Book |
| Language: | English |
| Published: |
[Place of publication not identified] :
[publisher not identified] ;
2001.
|
| Subjects: | |
| Online Access: | http://proxy.library.tamu.edu/login?url=http://proquest.umi.com/pqdweb?did=725906081&sid=1&Fmt=2&clientId=2945&RQT=309&VName=PQD |
| Summary: | This dissertation makes two primary contributions to the literature. The first contribution is to report on experiments that examine incentive problems in dominance solvable games. A novel matching protocol, T-Death, is presented. This protocol was used with two dominance solvable games. Game I was iterative dominance solvable; game S was strict dominance solvable. The results show that players more commonly adopted the payoff dominant symmetric play path in S. The terminal phase of S quickly became repeated play of the stage game equilibrium. Players confronted with game I had more difficulty establishing a play path that dominated the stage game payoffs. This difference in observed play paths is likely attributable to the difficulty players would have in distinguishing the other player's defection from an incorrect belief about an equilibrium assignment. The second primary contribution of this dissertation is to reexamine several aspects of the Experience-weighted Attraction learning model of Camerer & Ho (1998, 1999). In Camerer & Ho (1999), the authors enforce a restriction described as "The Law of Declining Effect." This restriction may be improper for games where the most recently observed strategy choices should most strongly guide a player's future choices. This dissertation reports on estimates of such games, and tests the effects of the restriction on the estimates. The results demonstrate that the law is likely inappropriate for these games and should be tested, not assumed. In Camerer & Ho (1998), one focus was on comparing the probability rules used to map the attractions to probability space over the strategies. Their specifications did not attempt to control for the different parameterization in one of the specifications. I examine the estimates after tightening up the underlying parameterization. The results suggest that there is not yet a clear choice of probability rule which should be used, if the log likelihood is the measure of appropriateness. Camerer & Ho (1998) also examined heterogeneity. Their estimation was over the entire sample and was a convex combination of two complete parameter sets. This dissertation reports on parameter estimates comparing session level data and data pooled across sessions. The results indicate there is generally a large improvement in fit gained by separating sessions. |
|---|---|
| Item Description: | Vita. "Major Subject: Economics". |
| Physical Description: | xiii, 171 leaves : illustrations ; 28 cm. Issued also on microfiche from University Microfilm Inc. |
| Bibliography: | Includes bibliographical references (leaves 112-116). |