Title :
Fast convergence in semi-anonymous potential games
Author :
Borowski, Holly ; Marden, Jason R. ; Frew, Eric W.
Author_Institution :
Dept. of Aerosp. Eng., Univ. of Colorado, Boulder, CO, USA
Abstract :
The log-linear learning algorithm has been extensively studied in game theoretic and distributed control literature. A central appeal of log-linear learning for distributed control is that it often guarantees agents´ behavior will converge in probability to the optimal configuration. However, one of its central issues is that the worst case convergence time can be prohibitively long, e.g., exponential in the number of players. We formalize a modified log-linear learning algorithm whose worst case convergence time is roughly linear in the number of players. We prove this characterization in semi-anonymous potential games with limited populations, i.e., in potential games where the agents´ utility functions can be expressed as a function of aggregate behavior within each population.
Keywords :
convergence; game theory; learning (artificial intelligence); multi-agent systems; probability; agents; aggregate behavior; convergence time; distributed control literature; game theoretic literature; log-linear learning algorithm; optimal configuration; probability; semianonymous potential games; utility functions; Convergence; Economics; Games; Markov processes; Nickel; Sociology; Statistics;
Conference_Titel :
Decision and Control (CDC), 2013 IEEE 52nd Annual Conference on
Conference_Location :
Firenze
Print_ISBN :
978-1-4673-5714-2
DOI :
10.1109/CDC.2013.6760242