| Izvleček: | <p>This thesis explores the intersection of causal and strategic reasoning, which are both fundamental concepts in AI, among many other disciplines. Historically, the absence of a unified formal framework accommodating both reasoning approaches has inhibited the formalisation of many important concepts. We offer a solution in the form of <em>(structural) causal games</em>, a framework which can be seen as extending Judea Pearl's causal hierarchy to the game-theoretic domain or as extending Daphne Koller and Brian Milch's multi-agent influence diagrams (MAIDs) to the causal domain. This dual extension paves the way for a holistic understanding and application of both causal and game-theoretic reasoning in complex decision-making scenarios.</p>
<p>This thesis makes several contributions toward our overall goal. First, we introduce mechanised MAIDs, an extension of MAIDs that enables one to make explicit the dependencies between agents' decision rules and the distributions governing the game. Second, we introduce subgames as well as several equilibrium refinements to MAIDs; in particular, our subgame perfectness refinement can rule out more non-credible threats than in extensive form games (EFGS) and can be exploited to find a Nash equilibrium in a MAID (in best case) exponentially faster than in an EFG. Third, we examine MAIDs in imperfect recall settings, where a Nash equilibrium in behavioural policies may not exist. We overcome this by showing how to solve MAIDs with forgetful and absent-minded agents using mixed policies and two types of correlated equilibrium. We also analyse the computational complexity of key decision problems in MAIDs, and explore tractable cases. Fourth, we present definitions of predictions, interventions, and counterfactuals in causal games and discuss the assumptions required for each. Finally, we showcase a variety of applications of causal games from designing safe AI systems to analysing economic policy interventions.</p>
|
|---|