Abstract
This paper develops a characteristics-based integral reinforcement learning(IRL) framework for the control of a nonlinear nonisothermal plug-flowreactor modeled as a convection--reaction hyperbolic distributed-parametersystem with distributed actuation. Exploiting the transport structure, themethod of characteristics yields an exact representation in terms of afamily of parameterized nonlinear ODE trajectories driven by re-indexeddistributed inputs. Building on this representation, an online IRLpolicy-iteration scheme is derived from an integral Bellman relation,enabling data-driven policy evaluation and improvement from measuredspatio-temporal trajectories without requiring explicit knowledge of thenonlinear drift dynamics. Monotonic improvement of the characteristicvalue functions is established for admissible policy updates, and aLyapunov functional induced by the characteristic value function is usedto establish a dissipation property for the closed-loop PDE. Numerical simulations demonstrate regulation toward a desiredsteady profile, robustness to bounded disturbances and moderatemodel uncertainties, and nominal closed-loop performancecomparable to offline LQR and IRL--LQR benchmarks.

This publication has 42 references indexed in Scilit: