Yoshinobu Kawahara
Yoshinobu Kawahara develops learning methods for changing systems, connecting predictions and control with the mathematical structure of dynamics. Kawahara coauthored Koopman Q-learning, which studied offline reinforcement learning through symmetries of dynamics. Another shared paper addressed discriminant dynamic-mode decomposition for labeled data collected across space and time. Kawahara directs RIKEN’s Dynamical Systems Learning Team. The team develops learning theories and algorithms for prediction and control, drawing on the structures behind data-generating processes.
- Now at
- RIKEN and University of Osaka
- Works
- 180
- Citations
- 2,449
- h-index
- 23
Publication numbers from OpenAlex, as of 14 October 2026.
Timeline
-
2022
Learning from the structure of change
Kawahara coauthored Koopman Q-learning, which studied offline reinforcement learning through symmetries of dynamics. Another shared paper addressed discriminant dynamic-mode decomposition for labeled data collected across space and time.
-
2023–2024
Stable models and useful symmetries
His work includes stable invariant models through Koopman spectra, neural dynamic-mode decomposition and symmetry-invariant Transformers for generalization in reinforcement learning. These methods connect mathematical properties of motion with learning algorithms.
-
2025
Stochastic and multi-agent dynamics
Kawahara coauthored work on learning stochastic nonlinear dynamics with embedded latent transfer operators. A separate paper studied counterfactual treatment outcomes over time in complex multi-agent settings.
-
As of October 2026
Dynamical systems learning
Kawahara directs RIKEN’s Dynamical Systems Learning Team. The team develops learning theories and algorithms for prediction and control, drawing on the structures behind data-generating processes.
Sources and credits
Photo: www.riken.jp; Yoshinobu Kawahara profile photograph; creator not stated. License: Not stated by source. Source: riken.jp. Cropped to a square and resized. Ask us to remove or replace this photo.