Authors:
(1) Avrim Blum, Toyota Technological Institute at Chicago, IL, USA;
(2) Melissa Dutz, Toyota Technological Institute at Chicago, IL, USA. Table of Links Abstract and 1 Introduction 2 Setting and 2.1 Models of behaviorally-biased opponents 3 Preliminaries and Intuition 4.1 Myopic Best Responder and 4.2 Gambler’s Fallacy Opponent 4.3 Win-Stay, Lose-Shift Opponent 4.4 Follow-the-Leader Opponent and 4.5 Highest Average Payoff Opponent 5 Generalizing 5.1 Other Behaviorally-Biased Strategies 5.2 Exploiting an Unknown Strategy from a Known Set of Strategies 6 Future Work and References A Appendix A.1 Win-Stay Lose-Shift Variant: Tie-Stay A.2 Follow-the-Leader Variant: Limited History A.3 Ellipsoid Mistake Bounds A.4 Highest Average Payoff Opponent 4.4 Follow-the-Leader Opponent Recall that the Follow-the-Leader opponent plays the best action in retrospect, defined as the action that would have achieved the highest payoff against our entire history of play. For this opponent, our strategy will be to learn a best response to each action, and then use the well-known ellipsoid algorithm to predict the opponent’s actions while playing best responses to the predicted actions. Using Ellipsoid for Prediction Note that while the bound on the number of losses and ties we incur is exponential, the runtime can be considered efficient since choosing which action to play in each round is efficient; this is an improvement over the simple general prediction algorithm we considered in section 3, which requires considering an exponential number of game matrices to choose which action to play in a single round. We also consider a limited-history variant of the Follow-the-Leader opponent below, against which we can achieve a polynomial bound on losses and ties. 4.4.1 Variant: Limited History 4.5 Highest Average Payoff Opponent The Highest Average Payoff opponent plays the action that has achieved the highest average payoff over the times they have played it. We discuss this opponent in A.4. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Avrim Blum, Toyota Technological Institute at Chicago, IL, USA; (2) Melissa Dutz, Toyota Technological Institute at Chicago, IL, USA. Authors: Authors: (1) Avrim Blum, Toyota Technological Institute at Chicago, IL, USA; (2) Melissa Dutz, Toyota Technological Institute at Chicago, IL, USA. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Setting and 2.1 Models of behaviorally-biased opponents 2 Setting and 2.1 Models of behaviorally-biased opponents 3 Preliminaries and Intuition 3 Preliminaries and Intuition 4.1 Myopic Best Responder and 4.2 Gambler’s Fallacy Opponent 4.1 Myopic Best Responder and 4.2 Gambler’s Fallacy Opponent 4.3 Win-Stay, Lose-Shift Opponent 4.3 Win-Stay, Lose-Shift Opponent 4.4 Follow-the-Leader Opponent and 4.5 Highest Average Payoff Opponent 4.4 Follow-the-Leader Opponent and 4.5 Highest Average Payoff Opponent 5 Generalizing 5.1 Other Behaviorally-Biased Strategies 5.1 Other Behaviorally-Biased Strategies 5.2 Exploiting an Unknown Strategy from a Known Set of Strategies 5.2 Exploiting an Unknown Strategy from a Known Set of Strategies 6 Future Work and References 6 Future Work and References A Appendix A.1 Win-Stay Lose-Shift Variant: Tie-Stay A.1 Win-Stay Lose-Shift Variant: Tie-Stay A.2 Follow-the-Leader Variant: Limited History A.2 Follow-the-Leader Variant: Limited History A.3 Ellipsoid Mistake Bounds A.3 Ellipsoid Mistake Bounds A.4 Highest Average Payoff Opponent A.4 Highest Average Payoff Opponent 4.4 Follow-the-Leader Opponent Recall that the Follow-the-Leader opponent plays the best action in retrospect, defined as the action that would have achieved the highest payoff against our entire history of play. For this opponent, our strategy will be to learn a best response to each action, and then use the well-known ellipsoid algorithm to predict the opponent’s actions while playing best responses to the predicted actions. Using Ellipsoid for Prediction Using Ellipsoid for Prediction Note that while the bound on the number of losses and ties we incur is exponential, the runtime can be considered efficient since choosing which action to play in each round is efficient; this is an improvement over the simple general prediction algorithm we considered in section 3, which requires considering an exponential number of game matrices to choose which action to play in a single round. We also consider a limited-history variant of the Follow-the-Leader opponent below, against which we can achieve a polynomial bound on losses and ties. 4.4.1 Variant: Limited History 4.5 Highest Average Payoff Opponent The Highest Average Payoff opponent plays the action that has achieved the highest average payoff over the times they have played it. We discuss this opponent in A.4. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

Ellipsoid Algorithms as a Tool Against Predictable Opponents

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

What AI Gets Wrong About Queer, Indigenous, and Racial Minorities

Understanding Bias-Driven Opponent Models in Competitive Gameplay

Ways to Counter Limited-History Opponents with Algorithmic Tools

Future Directions for Exploiting Behavioral Biases in Games

Broader Insights into Exploitable Strategies in Zero-Sum Games

How Behavioral Biases Shape Gameplay Without Payoff Visibility

What AI Gets Wrong About Queer, Indigenous, and Racial Minorities

Understanding Bias-Driven Opponent Models in Competitive Gameplay

Ways to Counter Limited-History Opponents with Algorithmic Tools

Future Directions for Exploiting Behavioral Biases in Games

Broader Insights into Exploitable Strategies in Zero-Sum Games

How Behavioral Biases Shape Gameplay Without Payoff Visibility

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps