Stochastic optimisation of vehicle suspension control systems via learning automata
This thesis considers the optimisation of vehicle suspension systems via a reinforcement learning technique The aim is to assess the potential of learning automata to learn 'optimum' control of suspension systems, which contain some active element under electronic control, without recourse to system models. Control optimisation tasks on full-active and senu-active suspension systems are used for the feasibility assessment and subsequent development of the learning automata technique. The quarter-vehicle simulation model, with ideal full-active suspension actuation, provides a well-known environment for initial studies applying classical discrete learning automata to learn the controller gains of a linear state-feedback controller. Learning automata are shown to be capable of acquiring near optimal controllers without any explicit knowledge of the suspension environment. However, the methodology has to be developed to allow safe on-line application. A moderator is introduced to prevent excessive suspension deviations as a result of possible unstable control actions applied during learning. A hardware trial is successfully implemented on a test vehicle fitted with semi-active suspension, excited by a hydraulic road simulation rig. During these initial studies some inherent weaknesses of the discrete automata are noted A discrete action set provides insufficient coverage of a continuous controller space so optima may be overlooked. Subsequent methods to increase the resolution of search lead to a forced convergence and hence an increased likelihood of local optima location. Th1s motivates the development of a new formulation of learning automaton, the CARLA, which exhibits a continuous action space and a reinforcement generalisation. The new method is compared w1th discrete automata on vanous stochastic function optimisatwn case stui1es, demonstrating that the new functionality of CARLA overcomes many of the identified shortcomings of discrete automata. Furthermore, CARLA shows a potential capability to learn in non-stationary environments. Repeatmg the earlier suspension tasks with CARLA applied, including an on-line hardware study, further demonstrates a performance gain over discrete automata Finally, a complex multi-goal learning task is considered A dynamic roll-control strategy is formulated based on the senu-active suspension hardware of the test vehicle. The CARLA is applied to the free parameters of this strategy and is seen to successfully synthesise improved roll-control over passive suspension.