Multi-Agent Multi-Objective Planning with Contextual Lexicographic Reward Preferences

Autonomous agents often operate in environments with multiple, conflicting objectives, where preference orderings vary with context. Existing multi-objective planning approaches assume a single preference ordering across the state space, making them unsuitable for context-dependent priorities. In multi-agent systems, this complexity increases as agents may operate under different contexts and must coordinate for safe joint operation. Current methods focus on static, single-agent settings or rely on centralized approaches that do not scale. My dissertation develops scalable, efficient techniques for multi-agent decision-making with context-dependent objective preferences. To that end, I developed a context-based planning framework for multi-objective settings, with theoretical guarantees for computing valid, cycle-free policies. I extended this framework to multi-agent systems, where agents complete tasks independently while mitigating negative side effects (NSEs) of joint actions. Future research will focus on contextual planning for cooperative multiagent systems under partial observability and non-stationarity, enabling lifelong autonomy in dynamic environments.