Abstract
Traditional synchronization of nonlinear biological oscillators often depends on fixed or manually tuned coupling parameters and this tends to be less adaptive when the physiological environment keeps changing over time. In contrast, existing reinforcement learning (RL) methods like Q-learning usually work with discrete action spaces, so they create kind of stepwise changes which can feel biologically unrealistic. Here we suggest a continuous-action Deep Deterministic Policy Gradient (DDPG) scheme, meant for adaptive identification and also control of a time-varying coupling strength in mixed or heterogeneous oscillator networks. In the setup, each oscillator is treated as a nonlinear dynamical system, while the states are linked through a single scalar, time dependent term k(t). That coefficient is updated in real time by the DDPG agent, so the coupling can track the changes instead of staying frozen. Our aim is to optimize two goals at the same time: minimize Mean Squared Error (MSE) and maximize the Phase-Locking Value (PLV) which together help lock the phases and keep a stable synchronization regime. We test the approach using five well known biological oscillator models: Circadian (Goodwin), Glycolytic (Sel'kov), Intracellular Calcium, NF-κB, and p53-Mdm2. The simulations show good phase entrainment online, plus data-driven discovery of the coupling term across all five systems. In terms of use cases, it could support cardiac rhythm regulation, neural coordination, and metabolic pathway organization.