Marginal HRL for Asymmetric Multi-Agent Systems

Problem

Classical MARL assumes homogeneous agents with fixed computational and communication capacity. Real systems (drones, robots, vehicles) differ in resources, observation limits, and safety requirements.

Hypothesis

Marginal Hierarchical Reinforcement Learning (M-HRL) can improve coordination by treating computation as a scarce resource and allocating a leader-controlled budget between safety and task layers across a coalition.

Possible Formalization

Lexicographic hierarchy: safety layer dominates reward pursuit; leader agent migrates computational priority based on marginal utility of control.

Possible Experiment

Small-scale asymmetric MARL simulation with variable agent compute budgets and safety-critical constraints.

Status

Concept note — early-stage.

See also the full concept post.