Behavioral analysis in the agent-based simulation of centipede games

The aim of this paper is to explain behavior of the human subjects in several laboratory experiments of centipede game which are different from the equilibrium solution. A centipede game is a dynamic game which multiple players make decision in prearranged order. Subgame perfect equilibrium (SPE) is a solution concept which is based on the concept of Nash equilibrium. SPE is well known solution as which predicts how the players play in most of dynamic games with the exception of few kinds of games. Some experimental results of centipede games which are different from the equilibrium are reported. Contrary to the theoretical assumption such that all players are rational and they can discriminate between slight difference of payoffs, it is thought of as human makind decision through trial-and-error process. This paper conducts simulation of the centipede game by using artificial adaptive agents, and shows behavioral features of the human subjects.

I. Introduction

A sequential game such that all players make decision in prearranged order is called a dynamic game or a extensive form game. Subgame perfect equilibrium is well known solution concept which predicts a strategy set of dynamic games. However, some experimental results of dynamic games, centipede games [14], [11], ultimatum bargaining games [3] and so forth, which the human subjects deviate from the equilibrium are reported. McKelvey and Palfrey [9], Fey et al. [4] and Nagel and Tang [12] conducted laboratory experiments of two person centipede games, and reported that the subgame perfect equilibrium are observed in some the experiments. Rapoport et al. [14] and Murphy et al. [11] conducted three person games. As the result, the subgame perfect equilibrium were rarely observed in the experiments of three person games.