My research focuses on theoretical and foundational advancements toward trustworthy sequential decision-making, with the vision of enabling the design of competent, autonomous systems that co-exist with humans, empower them, and are socially, ethically, and economically beneficial. Working at the intersection of learning and control, our group develops new models, algorithms, and theories that make autonomous agents capable of safe, efficient, and human-aware decision-making under uncertainty, with continual adaptation to new environments and tasks. The research problems we are interested in broadly relate to reinforcement learning, active perception, knowledge representation, online learning, control theory, and multi-agent systems. Some applications of interest for the designed solutions are robotics, disaster response, sustainability, healthcare, transportation, and entertainment. The three major directions we currently pursue, along with examples of our high-level methodological approach, are described below.
In this direction, we study active identification and gathering of actionable information from large-scale, multi-modal, and noisy data for control tasks. We formulate new classes of control-oriented information-gathering problems and analyze their computational and statistical complexity. We design task-oriented information utility metrics to quantify the importance of different information sources or parts of data given partial, uncertain knowledge about the environment and the task. We also develop efficient algorithms, accompanied by theoretical performance guarantees, to optimize for information utility given the resource constraints.
In this direction, we study complex planning and learning based on partial and evolving knowledge in real time. We design online learning algorithms with theoretical regret analyses for adaptive decision-making in a variety of settings, including structured bandits and reinforcement learning. We develop principled ways of incorporating structured side knowledge, such as automata-based task representations, probabilistic graphical models, causal models, and game-theoretic structures, to accelerate and improve sequential decision-making. In settings where such structures exist but are unknown a priori, we explore joint learning of these structured knowledge and optimal decision-making strategies.
In this direction, we study autonomous decision-making, for instance, for robots and more generally, AI systems, when interacting and collaborating with humans through different modalities. We design learning algorithms that utilize cognitive models to enable the autonomous agent to form a probabilistic understanding of human intent and preferences. We develop adaptive and safe decision-making strategies capable of incorporating the evolving knowledge based on human feedback. Furthermore, we investigate the interpretability and the explainability of the decisions in these human-AI interactive settings and analyze the potential trade-offs between these factors and the performance.