Meta-Value Learning: a General Framework for Learning with Learning Awareness

TL;DR

This paper introduces Meta-Value Learning (MeVa), a framework that evaluates joint policies based on their long-term optimization prospects, improving multi-agent gradient-based learning by considering future interactions.

Contribution

It proposes a novel meta-value approach that assesses policies over long-term horizons, avoiding explicit action space representation and REINFORCE estimators, enhancing multi-agent learning stability.

Findings

MeVa is consistent and far-sighted.

It outperforms prior methods in toy game scenarios.

The approach effectively captures long-term optimization prospects.

Abstract

Gradient-based learning in multi-agent systems is difficult because the gradient derives from a first-order model which does not account for the interaction between agents' learning processes. LOLA (arXiv:1709.04326) accounts for this by differentiating through one step of optimization. We propose to judge joint policies by their long-term prospects as measured by the meta-value, a discounted sum over the returns of future optimization iterates. We apply a form of Q-learning to the meta-game of optimization, in a way that avoids the need to explicitly represent the continuous action space of policy updates. The resulting method, MeVa, is consistent and far-sighted, and does not require REINFORCE estimators. We analyze the behavior of our method on a toy game and compare to prior work on repeated matrix games.