Optimal distributed learning for disturbance rejection in networked non‐linear games under unknown dynamics

Nonlinear Discrete-time System Identification without Persistence of Excitation: Finite-time Concurrent Learning Methods

“…However, the PE condition might be hard to achieve or even might not be feasible in some scenarios, especially in the context of on-line learning. Concurrent learning [2][3][4][5][6] has emerged as a promising paradigm in the direction that guarantees the exponential convergence of the estimated parameters to their optimal values with relaxing the strict assumption of the PE condition to some easy-to-check verifiable conditions on the richness of data. Concurrent learning technique benefits from recorded experienced data along with current data to replace the PE condition on the regressor with a rank condition on the memory stack of the regressor recorded data.…”

Assured learning‐enabled autonomy: A metacognitive reinforcement learning framework

“…In control community, several model-based and data driven RL-based feedback controllers have been presented for control of uncertain dynamical systems. [23][24][25][26][27][28] In these traditional RL-based controllers, the reinforcement signal feedback is derived through a fixed quadratic objective function. 29,30 A fixed reward or objective function, however, cannot guarantee achieving desired specifications across all circumstances.…”

Event‐triggered distributed H ∞ control of physically interconnected mobile Euler–Lagrange systems with slipping, skidding and dead zone

“…However, most of the works in ADP use known drift dynamics and the persistent excitation (PE) condition to guarantee the parameter convergence, for example, the recent work in [17]. Of course, the assumption of known drift dynamics makes the systems limited, and as shown in [18], PE is difficult to check online and makes the convergence process slow. Most recently, the concurrent learning technique of ADP is employed to relax the PE condition [19, 20] or is even combined with the PE condition to accelerate the convergence [21].…”

“…Unfortunately, external disturbance or input constraint is not included. ADP has also been employed to design optimal control schemes for systems with external disturbances and constrained inputs [17, 18, 22]. In the schemes, there are two policies of two players in a differential game, one for the optimal control law and the other for the worst disturbance law.…”

“…To mitigate the burden of computation bandwidth and data transfer costs, the controllers follow the event‐triggering mechanism, i.e. the control parameters are updated only when a new triggering condition holds. (ii) Extend the conventional theory of the zero‐sum differential graphical game of two players [17, 18, 20, 30] to design a new game of three players for approximating the event‐triggered control law, the event‐triggered dead‐zone disturbance law, and the time‐triggered slipping and skidding disturbance law. The new player who is responsible for rejecting the dead‐zone disturbance contributes to the optimisation of an performance index function.…”