GOV-REK: Governed Reward Engineering Kernels for Designing Robust Multi-Agent Reinforcement Learning Systems

For multi-agent reinforcement learning (MARL) systems, the problem task often involves massive problem-specific reward engineering effort. This effort is usually not directly transferable to other problems; worse, this problem is further exacerbated for sparse reward scenarios. We propose GOVerned Reward Engineering Kernels (GOV-REK), which dynamically assign reward distributions to agents in MARLs during the learning stage. We also introduce governance kernels, which exploit the underlying structure in either state or joint action space for assigning meaningful agent rewards. We demonstrate, using a Hyperband-like problem-agnostic algorithm, that this approach successfully learns to solve different MARL problems by iteratively exploring multiple reward models.