@article{Kadyrov2026, 
author = {Shirali Kadyrov and Ardak Kashkynbayev and Yershat Sapazhanov},
title = {Adaptive tracking control with Lyapunov regularization for data-calibrated Takagi–Sugeno fuzzy Markov jump biological systems},
year = {2026},
journal = {AIMS Mathematics},
volume = {11},
number = {7},
pages = {21167-21189},
keywords = {adaptive tracking control, reinforcement learning, fitted Q iteration, Lyapunov regularization, Takagi–Sugeno fuzzy systems, Markovian jump systems, tumour–immune dynamics, data-driven control},
url = {https://www.sciopen.com/article/10.3934/math.2026860},
doi = {10.3934/math.2026860},
abstract = {Cancer–tumour–immune dynamics are nonlinear, uncertain, and subject to abrupt biological or treatment-induced regime changes. This paper develops a data-calibrated adaptive tracking-control framework for Takagi–Sugeno (T–S) fuzzy Markov jump biological systems with unknown local dynamics and unknown transition probabilities. The proposed method combines mode-dependent fitted Q iteration with a Lyapunov-penalized policy update, so that the learned controller uses transition data while remaining close to a bounded stabilizing reference action. For the sampled discounted problem, we establish Bellman contraction, well-posedness of the fitted regression update, an approximate fitted Q error bound, a quantitative Lyapunov-penalization bound, a practical tracking guarantee for the Lyapunov-regularized policy, and positivity preservation under projected states and dimensionless inputs. Public tumour-volume data associated with the FIR atezolizumab non-small-cell lung cancer study are used only to calibrate plausible tumour-growth ranges; a trajectory-level holdout split is applied to avoid in-sample reporting. In the resulting active-treatment simulation regime, the proposed controller gives the best overall balanced performance among the compared strategies, which now include proximal policy optimization (PPO), soft actor–critic (SAC), nominal model predictive control (MPC), local tracking feedback, reference-action control, fixed scheduling, non-switching fitted Q iteration, and no control. In the representative single-seed run, the Lyapunov-regularized fitted Q iteration (LR-FQI) method achieves low weighted cost (   51.17), low tumour integral squared error (ISE) (   1.70), short threshold-exceedance time (   4 days), and moderate total dimensionless input. A Monte Carlo evaluation over 50 independent Markov and noise trajectories gives the lowest mean weighted cost for the proposed controller (   47.29  ±  19.91, mean    ± standard deviation) and a consistent threshold-exceedance time of    4.00  ±  0.00 days. Assumption 4 is verified numerically over 5000 sampled domain points.}
}