Tamer Başar

Research profile

Research summary

Tamer Başar is a researcher at the University of Illinois Urbana-Champaign; this profile is name-resolved from public OpenAlex bibliographic data. Topic counts in the snapshot place Game Theory and Applications first (11 tagged works), followed by stochastic processes and financial applications (5), optimization and variational analysis (4), adaptive dynamic programming control (4), and distributed control of multi-agent systems (3). Recent output splits exactly one-third each across three fronts: stochastic analysis for generative modeling, privacy-preserving decentralized learning, and reinforcement learning with stochastic control. Activity is the strongest measured axis (83/100): OpenAlex records 16 works in the last three years, and the Evolve catalog holds three 2026 arXiv preprints, each with this researcher as last author. The cataloged work is consistently theory-first: geometric-rate convergence proofs for privacy-preserving push-sum gossip learning, bi-level design of an agent's internal-state kernel with soft Q-learning convergence and zeroth-order outer-loop rates, and causal optimal-transport plus martingale-representation arguments yielding score-free regression losses for conditioning degenerate diffusion models. Evolve's verdicts repeatedly praise the rigor while flagging thin or qualitative experiments and, in the privacy paper, weak privacy accounting. Reproducibility averages 43/100 across the three assessed papers (two medium, one low), and collaboration is compact at five distinct coauthors.

Research keywords

Highlights

• Geometric-rate private gossip learning

  A last-authored 2026 arXiv preprint combining Tuning without Forgetting, robust differential privacy, and private push-sum consensus for single-sample gossip learning, with geometric-rate convergence proofs. Evolve finds the theory convincing but the privacy accounting less so; reproducibility readiness is rated medium.

• Bi-level design of adaptive agents

  A last-authored 2026 arXiv preprint on designing an agent's internal-state kernel for non-Markovian environments; it proves soft Q-learning convergence plus zeroth-order and Bayesian-optimization outer-loop rates. Evolve judges the theory sound but the experiments thin, and reproducibility readiness is rated low.

• Score-free losses for degenerate diffusions

  A last-authored 2026 arXiv preprint building causal optimal-transport and martingale-representation theory to condition degenerate diffusion models, yielding score-free regression losses. Evolve calls the theory rigorous while noting the experiments are qualitative; readiness is medium.

Timeline

1. 1994Earliest year of recorded work activity in the OpenAlex snapshot, with one work listed in countsByYear.
2. 2026Privacy Preserving Gossip Learning appears on arXiv: a last-authored paper combining Tuning without Forgetting, robust differential privacy, and private push-sum with geometric-rate convergence proofs.
3. 2026Adaptive Agent Design appears on arXiv: a last-authored bi-level design of an agent's internal-state kernel with soft Q-learning convergence and zeroth-order outer-loop rates.
4. 2026Conditioning Degenerate Diffusion Models appears on arXiv: a last-authored paper deriving score-free regression losses from causal optimal-transport and martingale-representation theory.

Papers in the Evolve catalog

• 2026-09-13

  Privacy Preserving Gossip Learning

• 2026-09-11

  Adaptive Agent Design

• 2026-09-03

  Conditioning Degenerate Diffusion Models

Frequent co-authors