We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Beyond variance reduction: Understanding the true impact of baselines on policy optimization

Abstract: Bandit and reinforcement learning (RL) problems can often be framed as optimization problems where the goal is to maximize average performance while having access only to stochastic estimates of the true gradient. Traditionally, stochastic optimization theory predicts that learning dynamics are governed by the curvature of the loss function and the noise of the gradient estimates. In this paper we demonstrate that this is not the case for bandit and RL problems. To allow our analysis to be interpreted in light of multi-step MDPs, we focus on techniques derived from stochastic optimization principles (e.g., natural policy gradient and EXP3) and we show that some standard assumptions from optimization theory are violated in these problems. We present theoretical results showing that, at least for bandit problems, curvature and noise are not sufficient to explain the learning dynamics and that seemingly innocuous choices like the baseline can determine whether an algorithm converges. These theoretical findings match our empirical evaluation, which we extend to multi-state MDPs.
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:2008.13773 [cs.LG]
  (or arXiv:2008.13773v3 [cs.LG] for this version)

Submission history

From: Wesley Chung [view email]
[v1] Mon, 31 Aug 2020 17:52:09 GMT (6819kb,D)
[v2] Mon, 7 Dec 2020 18:13:05 GMT (14871kb,D)
[v3] Fri, 19 Feb 2021 18:10:59 GMT (13704kb,D)

Link back to: arXiv, form interface, contact.