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: Modularity in Reinforcement Learning via Algorithmic Independence in Credit Assignment

Abstract: Many transfer problems require re-using previously optimal decisions for solving new tasks, which suggests the need for learning algorithms that can modify the mechanisms for choosing certain actions independently of those for choosing others. However, there is currently no formalism nor theory for how to achieve this kind of modular credit assignment. To answer this question, we define modular credit assignment as a constraint on minimizing the algorithmic mutual information among feedback signals for different decisions. We introduce what we call the modularity criterion for testing whether a learning algorithm satisfies this constraint by performing causal analysis on the algorithm itself. We generalize the recently proposed societal decision-making framework as a more granular formalism than the Markov decision process to prove that for decision sequences that do not contain cycles, certain single-step temporal difference action-value methods meet this criterion while all policy-gradient methods do not. Empirical evidence suggests that such action-value methods are more sample efficient than policy-gradient methods on transfer problems that require only sparse changes to a sequence of previously optimal decisions.
Comments: Long Presentation at the Thirty-eighth International Conference on Machine Learning (ICML) 2021. 21 pages, 11 figures. v2: updated acknowledgments. v3: clarified that the internal function nodes of the credit assignment mechanism are not considered O(1)
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Information Theory (cs.IT); Neural and Evolutionary Computing (cs.NE); Machine Learning (stat.ML)
Cite as: arXiv:2106.14993 [cs.LG]
  (or arXiv:2106.14993v3 [cs.LG] for this version)

Submission history

From: Michael Chang [view email]
[v1] Mon, 28 Jun 2021 21:29:13 GMT (1004kb,D)
[v2] Sat, 3 Jul 2021 21:42:55 GMT (1004kb,D)
[v3] Wed, 21 Jul 2021 17:07:10 GMT (1005kb,D)

Link back to: arXiv, form interface, contact.