TY - GEN
T1 - Can agents acquire human-like behaviors in a sequential bargaining game? - Comparison of roth's and Q-learning agents
AU - Takadama, Keiki
AU - Kawai, Tetsuro
AU - Koyama, Yuhsuke
PY - 2007
Y1 - 2007
N2 - This paper addresses agent modeling in multiagent-based simulation (MABS) to explore agents who can reproduce human-like behaviors in the sequential bargaining game, which is more difficult to be reproduced than in the ultimate game (i.e., one time bargaining game). For this purpose, we focus on the Roth's learning agents who can reproduce human-like behaviors in several simple examples including the ultimate game, and compare simulation results of Roth's learning agents and Q-learning agents in the sequential bargaining game. Intensive simulations have revealed the following implications: (1) Roth's basic and three parameter reinforcement learning agents with any type of three action selections (i.e., e-greed, roulette, and Boltzmann distribution selections) can neither learn consistent behaviors nor acquire sequential negotiation in sequential bargaining game; and (2) Q-learning agents with any type of three action selections, on the other hand, can learn consistent behaviors and acquire sequential negotiation in the same game. However, Q-learning agents cannot reproduce the decreasing trend found in subject experiments.
AB - This paper addresses agent modeling in multiagent-based simulation (MABS) to explore agents who can reproduce human-like behaviors in the sequential bargaining game, which is more difficult to be reproduced than in the ultimate game (i.e., one time bargaining game). For this purpose, we focus on the Roth's learning agents who can reproduce human-like behaviors in several simple examples including the ultimate game, and compare simulation results of Roth's learning agents and Q-learning agents in the sequential bargaining game. Intensive simulations have revealed the following implications: (1) Roth's basic and three parameter reinforcement learning agents with any type of three action selections (i.e., e-greed, roulette, and Boltzmann distribution selections) can neither learn consistent behaviors nor acquire sequential negotiation in sequential bargaining game; and (2) Q-learning agents with any type of three action selections, on the other hand, can learn consistent behaviors and acquire sequential negotiation in the same game. However, Q-learning agents cannot reproduce the decreasing trend found in subject experiments.
KW - Agent modeling
KW - Agent-based simulation
KW - Human-like behaviors
KW - Reinforcement learning
KW - Sequential bargaining game
UR - http://www.scopus.com/inward/record.url?scp=37349063167&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=37349063167&partnerID=8YFLogxK
U2 - 10.1007/978-3-540-76539-4_12
DO - 10.1007/978-3-540-76539-4_12
M3 - Conference contribution
AN - SCOPUS:37349063167
SN - 9783540765363
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 156
EP - 171
BT - Multi-Agent-Based Simulation VII - International Workshop, MABS 2006, Revised and Invited Papers
PB - Springer Verlag
T2 - 7th International Workshop on Multi-Agent-Based Simulation, MABS 2006
Y2 - 8 May 2007 through 8 May 2007
ER -