Cooperative Multi-Agent Alignment via Boolean Task Algebras and Team Morality Chains

Abstract

Cooperative multi-agent reinforcement learning (MARL) offers a principled route to deploying teams of autonomous agents, but standard scalar-reward optimisation can produce misaligned behaviour in safety-critical settings. This PhD research studies cooperative multi-agent alignment, decomposed into intention alignment (faithful execution of specified tasks) and value alignment (strict adherence to priority-ordered normative constraints). For intention alignment, I develop agent-level compositional task specification via a cooperative extension of Boolean Task Algebras, paired with goal-oriented learning to support zero-shot generalisation across tasks. For value alignment, I build on MoralityGym and morality chains and develop a lexicographical reinforcement learning approach based on principled scalarisation to enforce team-level moral priorities. I outline how these components integrate by treating task reward as the lowest-priority objective within a Team Morality Chain, yielding cooperative policies that execute intended tasks subject to non-negotiable safety constraints.