Modeling Rationality: Toward Better Performance Against Unknown Agents in Sequential Games.

Area of Science:

• Artificial Intelligence
• Game Theory
• Autonomous Systems

Background:

• Opponent modeling is crucial for autonomous agents in strategic interactions.
• Existing methods often require extensive interaction history, which is frequently unrealistic.
• Imperfect information games present unique challenges for opponent modeling due to hidden states.

Purpose of the Study:

• To develop a novel opponent modeling method for games with imperfect information.
• To address the limitations of existing methods that rely on large interaction datasets.
• To improve the accuracy and adaptability of autonomous agents in strategic scenarios.

Main Methods:

• Introduced a Rationality-Consistent Opponent Modeling (ROM) approach.
• Utilized game-theoretical concepts of rationality consistency across disjoint information sets.
• Incorporated an online heuristic adaptation for reduced computational cost.

Main Results:

• ROM accurately infers opponent strategies even with insufficient observation history.
• Demonstrated superior performance and prediction accuracy compared to other methods in grid world and poker games.
• Heuristic adaptation significantly reduced the computational time cost of the opponent model.

Conclusions:

• ROM effectively models opponents in imperfect information games with limited data.
• The method enhances agent adaptability and prediction accuracy against diverse opponents.
• The heuristic adaptation makes ROM computationally efficient for real-time applications.