Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Collaborative Value Function Estimation Under Model Mismatch: A Federated Temporal Difference Analysis
Stockholm University, Faculty of Social Sciences, Department of Computer and Systems Sciences.ORCID iD: 0000-0003-4884-4600
King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia.
King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia.
Stockholm University, Faculty of Social Sciences, Department of Computer and Systems Sciences.ORCID iD: 0000-0002-6617-8683
Number of Authors: 42025 (English)In: Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2025, Porto, Portugal, September 15–19, 2025, Proceedings, Part VI / [ed] Rita P. Ribeiro; Bernhard Pfahringer; Nathalie Japkowicz; Pedro Larrañaga; Alípio M. Jorge; Carlos Soares; Pedro H. Abreu; João Gama, Springer , 2025, p. 41-58Conference paper, Published paper (Refereed)
Abstract [en]

Federated reinforcement learning (FedRL) enables collaborative learning while preserving data privacy by preventing direct data exchange between agents. However, many existing FedRL algorithms assume that all agents operate in identical environments, which is often unrealistic. In real-world applications, such as multi-robot teams, crowdsourced systems, and large-scale sensor networks, each agent may experience slightly different transition dynamics, leading to inherent model mismatches. In this paper, we first establish linear convergence guarantees for single-agent temporal difference learning (TD(0)) in policy evaluation and demonstrate that under a perturbed environment, the agent suffers a systematic bias that prevents accurate estimation of the true value function. This result holds under both i.i.d. and Markovian sampling regimes. We then extend our analysis to the federated TD(0) (FedTD(0)) setting, where multiple agents, each interacting with its own perturbed environment, periodically share value estimates to collaboratively approximate the true value function of a common underlying model. Our theoretical results indicate the impact of model mismatch, network connectivity, and mixing behavior on the convergence of FedTD(0). Empirical experiments corroborate our theoretical gains, highlighting that even moderate levels of information sharing significantly mitigate environment-specific errors.

Place, publisher, year, edition, pages
Springer , 2025. p. 41-58
Series
Lecture Notes in Computer Science (LNCS), ISSN 0302-9743, E-ISSN 1611-3349 ; 16018
Keywords [en]
Federated Reinforcement Learning, Model Mismatch in Reinforcement Learning, Temporal Difference Learning, Policy Evaluation
National Category
Computer Sciences
Research subject
Computer and Systems Sciences
Identifiers
URN: urn:nbn:se:su:diva-248239DOI: 10.1007/978-3-032-06106-5_3Scopus ID: 2-s2.0-105020022174ISBN: 978-3-032-06106-5 (electronic)ISBN: 978-3-032-06105-8 (print)OAI: oai:DiVA.org:su-248239DiVA, id: diva2:2007752
Conference
European Conference, ECML PKDD 2025, Porto, Portugal, September 15–19, 2025.
Available from: 2025-10-20 Created: 2025-10-20 Last updated: 2026-07-16Bibliographically approved
In thesis
1. Toward Scalable, Robust, and Efficient Reinforcement Learning: Theoretical Foundations and Algorithms for Sample-Efficient, Communication-Aware, and Heterogeneity-Robust RL
Open this publication in new window or tab >>Toward Scalable, Robust, and Efficient Reinforcement Learning: Theoretical Foundations and Algorithms for Sample-Efficient, Communication-Aware, and Heterogeneity-Robust RL
2026 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

Reinforcement learning (RL) holds the promise of autonomous decision-making in robotics, energy management, and personalized healthcare, yet three practical obstacles curb its impact. A solitary agent typically needs an impractically large number of interactions before useful behavior emerges; distributing the workload across many learners shortens the wall-clock time but burdens the network with costly model exchanges, and once experience is gathered on diverse devices, differences in local data and dynamics erode the guarantees enjoyed by standard federated algorithms.

This thesis offers a unified response that advances sample efficiency, communication awareness, and heterogeneity robustness in tandem. For single-agent settings, it introduces two complementary forms of lightweight guidance: an automatically annealed curriculum of auxiliary goals that accelerates exploration in sparse-or-composite-reward tasks, and a decaying surrogate loss that soft-injects imperfect heuristic controllers during early training. In distributed scenarios, a new convergence analysis proves that momentum-based optimization remains stable even when updates are aggressively compressed, leading to a federated Q-learning framework that preserves linear progress while transmitting an order-of-magnitude fewer bits. Finally, the work establishes convergence bounds for federated optimization without tying step-size choices to explicit data similarity constants and extends them to collaborative value estimation when each learner operates in a perturbed environment, revealing precisely how mismatch and communication frequency shape residual error. Experiments across tabular and high-dimensional decision-making domains confirm that interaction demands plummet, bandwidth usage drops sharply, and learning degrades gracefully under severe non-IID conditions. Collectively, the six papers assembled here chart a coherent route toward sample-efficient, communication-constrained, and heterogeneity-robust RL, advancing the feasibility of real-world deployment on resource-limited, privacy-sensitive platforms.

Abstract [sv]

Förstärkningsinlärning (reinforcement learning, RL) utlovar självgående beslutsfattande inom områden som robotik, energihantering och personcentrerad sjukvård. Tre praktiska hinder begränsar dock teknikens genomslag. En ensam agent behöver ofta ett orimligt stort antal interaktioner innan användbart beteende uppstår. När flera agenter tränas parallellt blir visserligen väggklocktiden kortare, men nätverket belastas i stället av omfattande modellutbyten. Dessutom undergrävs teoretiska garantier när lärandet sker på heterogena enheter med skilda datakällor och dynamik.

Avhandlingen presenterar ett samlat angreppssätt som samtidigt höjer proveffektiviteten, minskar kommunikationsbehovet och ökar robustheten mot heterogenitet. För enskilda agenter introduceras två kompletterande, lättviktiga styrmekanismer: dels ett automatiskt avtonande målkurrikulum som påskyndar utforskningen i glesa eller sammansatta belöningsmiljöer, dels en avtagande surrogatförlust som injicerar ofullkomliga heuristiska kontrollregler under de tidiga träningsfaserna. I distribuerade scenarier visar en ny konvergensanalys att momentumbaserad optimering förblir stabil även när uppdateringar kraftigt komprimeras. Resultatet är ett federerat Q-lärningsramverk som bibehåller linjär inlärningstakt samtidigt som det sänder en storleksordning färre bitar. Vidare härleds konvergensgränser för federerad optimering utan att steglängden behöver kopplas till någon datasimilaritetskonstant, och dessa utvidgas till kollaborativ värdeestimering när varje agent verkar i en perturberad miljö, vilket klargör hur modellskillnader och synkroniseringsfrekvens påverkar kvarstående fel. Experiment på såväl tabellära som högdimensionella beslutsproblem bekräftar analysen: interaktionskraven sjunker drastiskt, bandbreddsanvändningen minskar kraftigt och inlärningen degraderas kontrollerat även under starkt icke-IID-betingade förhållanden. Sammantaget utstakar de sex inkluderade studierna en tydlig väg mot proveffektiv, kommunikationssnål och heterogenitetstålig RL och för därmed tekniken ett steg närmare verklig tillämpning på resursbegränsade och integritetskänsliga plattformar.

Place, publisher, year, edition, pages
Stockholm: Department of Computer and Systems Sciences, Stockholm University, 2026. p. 104
Series
Report Series / Department of Computer & Systems Sciences, ISSN 1101-8526 ; 26-008
Keywords
Artificial Intelligence, Machine Learning, Reinforcement Learning, Federated Learning, Federated Reinforcement Learning
National Category
Computer Systems Computer Vision and Learning Systems Communication Systems Robotics and automation Control Engineering
Research subject
Computer and Systems Sciences
Identifiers
urn:nbn:se:su:diva-257712 (URN)978-91-8107-680-6 (ISBN)978-91-8107-681-3 (ISBN)
Public defence
2026-10-16, L50, Borgarfjordsgatan 12, Kista, Stockholm, 13:00 (English)
Opponent
Supervisors
Available from: 2026-09-23 Created: 2026-07-16 Last updated: 2026-08-27Bibliographically approved

Open Access in DiVA

No full text in DiVA

Other links

Publisher's full textScopus

Authority records

Beikmohammadi, AliMagnússon, Sindri

Search in DiVA

By author/editor
Beikmohammadi, AliMagnússon, Sindri
By organisation
Department of Computer and Systems Sciences
Computer Sciences

Search outside of DiVA

GoogleGoogle Scholar

doi
isbn
urn-nbn

Altmetric score

doi
isbn
urn-nbn
Total: 48 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf