Trending

Bookmarks

Keep Up With Latest Trending Papers. Computer Science, AI and more.Subscribe

Machine Learning

Multiagent Systems

Artificial Intelligence

MA-Trace: An On-Policy Actor-Critic Algorithm for Multi-Agent Reinforcement Learning

Off-Policy Correction For Multi-Agent Reinforcement Learning

We propose a new on-policy actor-critic algorithm for multi-agent reinforcement learning (marl).

Our algorithm, named ma-trace, is a generalization of the popular actor-critic algorithm (v-trace) to the multi-agent setting.

It utilizes importance sampling as an off-policy correction method, which allows distributing the computations with no impact on the quality of training.

We evaluate the algorithm extensively on the starcraft multi-agent challenge, a standard benchmark for multi-agentalgorithms, and show that it achieves high performance on all its tasks and exceeds state-of-the-art results on some of them.

Authors

Michał Zawalski, Błażej Osiński, Henryk Michalewski, Piotr Miłoś

Related Topics

Importance sampling

Multi-agent reinforcement learning (marl)

Multi-agent algorithms

On-policy actor-critic algorithm

Off-policy correction method

Read the Paper

Read Paper

Trending

Bookmarks

◐ Latest Activity

More

◐ Recommended Follows

More

MA-Trace: An On-Policy Actor-Critic Algorithm for Multi-Agent Reinforcement Learning - 42Papers

Read Related Papers

Learning Interactive Driving Policies via Data-driven Simulation

Memory-Efficient Episodic Control Reinforcement Learning with Dynamic Online k-means

Learning State Representations via Retracing in Reinforcement Learning

Adaptively Calibrated Critic Estimates for Deep Reinforcement Learning

Multi-Agent Reinforcement Learning for Dynamic Energy Storage Systems

Decentralized Multi-Armed Bandit Can Outperform Classic Upper Confidence Bound

Offline Reinforcement Learning: Fundamental Barriers for Value Function Approximation

Gradient Temporal Difference with Momentum: Stability and Convergence

Episodic Multi-agent Reinforcement Learning with Curiosity-Driven Exploration

UMBRELLA: Uncertainty-Aware Model-Based Offline Reinforcement Learning Leveraging Planning