March 26, 2010
Playing the Lemonade Game: Metacognition in Game Simulations.
A few details about our work on the Lemonade Stand Game leading up to our entry in the Lemonade competition at the end of 2009 can be found in our latest paper, presented at this weeks Behavior Representation in Modeling & Simulation conference. The competition was a tournament between agents, playing out many strategies that all tries to succeed to win lemonade market share on the beaches of a fantasy island.
Abstract: The Lemonade Game is a three-player game in which players have to pick locations on a circular board, which are as far away as possible from those chosen independently by other players. Players may observe other player's moves and infer their strategies. The game was studied using a competition of cognitively motivated agents, which inherit properties of adaptivity and stochasticity from human memory and decision-making, and simplistic, yet effective agents implementing fixed strategies. We argue that metacognition is the unique attribute that allows sophisticated agents to adapt to unforeseen conditions, cooperators and competitors.
D. Reitter, I. Juvina, A. Stocco, and C. Lebiere. Resistance is futile: Winning lemonade market share through metacognitive reasoning in a three-agent cooperative game. In Proceedings of the 19th Behavior Representation in Modeling & Simulation (BRIMS), Charleston, SC, 2010.
Metacognition was also explored in my cognitive model submitted to the Dynamic Stocks&Flows Competition last year. There, the model did just what I think human participants in a psychological experiment did: try different strategies to solve a problem, and pick the one that performed best most recently. Weighing recency and frequency of experiences or pieces of knowledge is a common theme in such models; more research is needed to find out how exactly people combine their experiences to make (mostly) good decisions. The model won the first prize in the challenge. Still, the model does not explain enough of the variance observed in the empirical data. A look back to determine what we have learned from the DSF challenge was presented in a panel discussion at this year's Behavior Representation in Modeling & Simulation conference.
Posted by dr at March 26, 2010 3:32 PM