Research

Central Question: In a two-player adversarial interaction with asymmetric information, how can the player with superior information gain advantage over the opponent by employing deception?

Idea: In real-world, agents need the capability of making decisions with incomplete information. Incomplete information refers to situations in which the agent or its environment is unaware of the other agent's action capabilities, objectives (payoffs), the transition dynamics or the other agent's knowledge. On becoming aware of the misinformation of its opponent, the agent can leverage it to deceive the opponent and achieve a better outcome.

Contributions: In my research, I have developed hypergame theory for games on graphs (a.k.a., omega-regular hypergame) to model and (qualitatively) analyze two-player adversarial interactions under various classes of incomplete information. This includes definition of various solution concepts and design of algorithms to synthesize correct-by-construction deceptive strategies from linear temporal logic (LTL) specifications.

Central Question: Given an incomplete (i.e., partial) preference over a set of linear temporal logic (LTL) formulas expressing temporal goals for an agent, how to synthesize a strategy that achieves the most preferred goal while reasoning about the uncertainties in the stochastic environment?

Idea: An cognitively realistic autonomous system must be able to simultaneously reason about multiple objectives and achieve the best outcome. A key challenge here is to be able reason with incomplete preferences. Incomplete preferences may be a result of inescapability; wherein the agent must make a decision under time and memory constraints, or incomplete information; wherein the agent does not know the user's complete preferences, or incomparability; i.e., when the outcomes are fundamentally incomparable.

Contributions: In my research, I am developing an automata-theoretic approach by defining a new language that can express preferences over linear temporal logic (LTL) objectives. I have shown that such a language has an automata-theoretic representation which can be used to synthesize strategies that are guaranteed to achieve the best outcomes under qualitative and quantitative solution concepts.

An interesting consequence of preference-based planning in Markov decision processes is that agents can exhibit opportunistic behaviors.

Central Question: How to plan qualitatively (i.e., synthesize sure, almost-sure, positive winning strategies) in two-player partially observable stochastic games (POSG) in which the adversary can attack the observation function of the first player?

Contributions: In addition to incomplete information, an autonomous agent must be able to reason with imperfect information (partial observation). In this collaborative research, we are developing algorithms to synthesize a strategy using which the agent can achieve its objective even under adversarial attacks. We study two cases: when agent is unaware that it is under attack, but assumes that sensor failures are probabilistic occurrence. And, when the agent is aware of adversary's presence and capabilities.