Stabilising Experience Replay for Deep Multi-Agent Reinforcement Learning

Authors: Jakob Foerster, Nantas Nardelli, Gregory Farquhar, Triantafyllos Afouras, Philip H. S. Torr, Pushmeet Kohli, Shimon Whiteson

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Results on a challenging decentralised variant of Star Craft unit micromanagement confirm that these methods enable the successful combination of experience replay with multi-agent RL.
Researcher Affiliation Collaboration 1University of Oxford, Oxford, United Kingdom 2Microsoft Research, Redmond, USA.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper mentions using Torch Craft, but does not state that the code for the methodology described in this paper is open-source or provide a link to it.
Open Datasets No The paper uses the Star Craft unit micromanagement environment but does not provide concrete access information for a publicly available or open dataset.
Dataset Splits No The paper describes training on sampled episodes from an experience replay memory but does not provide specific dataset split information (e.g., percentages or counts for training, validation, and test sets).
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions 'Torch7' and 'Torch Craft' but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes We linearly anneal ϵ from 1.0 to 0.02 over 1500 episodes, and train the network for emax = 2500 training episodes. In the standard training loop, we collect a single episode and add it to the replay memory at each training step. We sample batches of 30 n episodes uniformly from the replay memory and train on fully unrolled episodes. In order to reduce the variance of the multi-agent importance weights, we clip them to the interval [0.01, 2].