Reactive bandits with attitude

Pedro Ortega, Kee-Eung Kim, Daniel Lee
; Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics, PMLR 38:726-734, 2015.

Abstract

We consider a general class of K-armed bandits that adapt to the actions of the player. A single continuous parameter characterizes the “attitude” of the bandit, ranging from stochastic to cooperative or to fully adversarial in nature. The player seeks to maximize the expected return from the adaptive bandit, and the associated optimization problem is related to the free energy of a statistical mechanical system under an external field. When the underlying stochastic distribution is Gaussian, we derive an analytic solution for the long run optimal player strategy for different regimes of the bandit. In the fully adversarial limit, this solution is equivalent to the Nash equilibrium of a two-player, zero-sum semi-infinite game. We show how optimal strategies can be learned from sequential draws and reward observations in these adaptive bandits using Bayesian filtering and Thompson sampling. Results show the qualitative difference in policy pseudo-regret between our proposed strategy and other well-known bandit algorithms.

Cite this Paper


BibTeX
@InProceedings{pmlr-v38-ortega15, title = {{Reactive bandits with attitude}}, author = {Pedro Ortega and Kee-Eung Kim and Daniel Lee}, booktitle = {Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics}, pages = {726--734}, year = {2015}, editor = {Guy Lebanon and S. V. N. Vishwanathan}, volume = {38}, series = {Proceedings of Machine Learning Research}, address = {San Diego, California, USA}, month = {09--12 May}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v38/ortega15.pdf}, url = {http://proceedings.mlr.press/v38/ortega15.html}, abstract = {We consider a general class of K-armed bandits that adapt to the actions of the player. A single continuous parameter characterizes the “attitude” of the bandit, ranging from stochastic to cooperative or to fully adversarial in nature. The player seeks to maximize the expected return from the adaptive bandit, and the associated optimization problem is related to the free energy of a statistical mechanical system under an external field. When the underlying stochastic distribution is Gaussian, we derive an analytic solution for the long run optimal player strategy for different regimes of the bandit. In the fully adversarial limit, this solution is equivalent to the Nash equilibrium of a two-player, zero-sum semi-infinite game. We show how optimal strategies can be learned from sequential draws and reward observations in these adaptive bandits using Bayesian filtering and Thompson sampling. Results show the qualitative difference in policy pseudo-regret between our proposed strategy and other well-known bandit algorithms.} }
Endnote
%0 Conference Paper %T Reactive bandits with attitude %A Pedro Ortega %A Kee-Eung Kim %A Daniel Lee %B Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2015 %E Guy Lebanon %E S. V. N. Vishwanathan %F pmlr-v38-ortega15 %I PMLR %J Proceedings of Machine Learning Research %P 726--734 %U http://proceedings.mlr.press %V 38 %W PMLR %X We consider a general class of K-armed bandits that adapt to the actions of the player. A single continuous parameter characterizes the “attitude” of the bandit, ranging from stochastic to cooperative or to fully adversarial in nature. The player seeks to maximize the expected return from the adaptive bandit, and the associated optimization problem is related to the free energy of a statistical mechanical system under an external field. When the underlying stochastic distribution is Gaussian, we derive an analytic solution for the long run optimal player strategy for different regimes of the bandit. In the fully adversarial limit, this solution is equivalent to the Nash equilibrium of a two-player, zero-sum semi-infinite game. We show how optimal strategies can be learned from sequential draws and reward observations in these adaptive bandits using Bayesian filtering and Thompson sampling. Results show the qualitative difference in policy pseudo-regret between our proposed strategy and other well-known bandit algorithms.
RIS
TY - CPAPER TI - Reactive bandits with attitude AU - Pedro Ortega AU - Kee-Eung Kim AU - Daniel Lee BT - Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics PY - 2015/02/21 DA - 2015/02/21 ED - Guy Lebanon ED - S. V. N. Vishwanathan ID - pmlr-v38-ortega15 PB - PMLR SP - 726 DP - PMLR EP - 734 L1 - http://proceedings.mlr.press/v38/ortega15.pdf UR - http://proceedings.mlr.press/v38/ortega15.html AB - We consider a general class of K-armed bandits that adapt to the actions of the player. A single continuous parameter characterizes the “attitude” of the bandit, ranging from stochastic to cooperative or to fully adversarial in nature. The player seeks to maximize the expected return from the adaptive bandit, and the associated optimization problem is related to the free energy of a statistical mechanical system under an external field. When the underlying stochastic distribution is Gaussian, we derive an analytic solution for the long run optimal player strategy for different regimes of the bandit. In the fully adversarial limit, this solution is equivalent to the Nash equilibrium of a two-player, zero-sum semi-infinite game. We show how optimal strategies can be learned from sequential draws and reward observations in these adaptive bandits using Bayesian filtering and Thompson sampling. Results show the qualitative difference in policy pseudo-regret between our proposed strategy and other well-known bandit algorithms. ER -
APA
Ortega, P., Kim, K. & Lee, D.. (2015). Reactive bandits with attitude. Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics, in PMLR 38:726-734

Related Material