Ensemble Value Functions for Efficient Exploration in Multi-Agent Reinforcement Learning
Abstract: Existing value-based algorithms for cooperative multi-agent reinforcement learning (MARL) commonly rely on random exploration, such as $\epsilon$-greedy, to explore the environment. However, such exploration is inefficient at finding effective joint actions in states that require cooperation of multiple agents. In this work, we propose ensemble value functions for multi-agent exploration (EMAX), a general framework to seamlessly extend value-based MARL algorithms with ensembles of value functions. EMAX leverages the ensemble of value functions to guide the exploration of agents, stabilises their optimisation, and makes their policies more robust to miscoordination. These benefits are achieved by using a combination of three techniques. (1) EMAX uses the uncertainty of value estimates across the ensemble in a UCB policy to guide the exploration. This exploration policy focuses on parts of the environment which require cooperation across agents and, thus, enables agents to more efficiently learn how to cooperate. (2) During the optimisation, EMAX computes target values as average value estimates across the ensemble. These targets exhibit lower variance compared to commonly applied target networks, leading to significant benefits in MARL which commonly suffers from high variance caused by the exploration and non-stationary policies of other agents. (3) During evaluation, EMAX selects actions following a majority vote across the ensemble, which reduces the likelihood of selecting sub-optimal actions. We instantiate three value-based MARL algorithms with EMAX, independent DQN, VDN and QMIX, and evaluate them in 21 tasks across four environments. Using ensembles of five value functions, EMAX improves sample efficiency and final evaluation returns of these algorithms by 60%, 47%, and 539%, respectively, averaged across 21 tasks.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Pablo_Samuel_Castro1
Submission Number: 2515
Loading