English

One-Shot Averaging for Distributed TD($\lambda$) Under Markov Sampling

Machine Learning 2024-06-04 v2 Distributed, Parallel, and Cluster Computing

Abstract

We consider a distributed setup for reinforcement learning, where each agent has a copy of the same Markov Decision Process but transitions are sampled from the corresponding Markov chain independently by each agent. We show that in this setting, we can achieve a linear speedup for TD(λ\lambda), a family of popular methods for policy evaluation, in the sense that NN agents can evaluate a policy NN times faster provided the target accuracy is small enough. Notably, this speedup is achieved by ``one shot averaging,'' a procedure where the agents run TD(λ\lambda) with Markov sampling independently and only average their results after the final step. This significantly reduces the amount of communication required to achieve a linear speedup relative to previous work.

Keywords

Cite

@article{arxiv.2403.08896,
  title  = {One-Shot Averaging for Distributed TD($\lambda$) Under Markov Sampling},
  author = {Haoxing Tian and Ioannis Ch. Paschalidis and Alex Olshevsky},
  journal= {arXiv preprint arXiv:2403.08896},
  year   = {2024}
}