Vol.11, No.1 (1998) 1-24
Abstract: In this paper several nonparametric supervised machine learning (ML) techniques for automatic designing of rules for a rule-based control (RBC) of functional electrical stimulation (FES) assisted human walking are described. The application of an artificial neural network with radial basis functions, which works the best for the required pattern matching, is presented. ML can describe behavior of the system under certain conditions by performing spatio-temporal mapping of input-output variables and store it in appropriate form for use in real-time control. This approach is applicable whenever there is a skilled repetitive action or a process involving human or natural control. RBC relies on a finite state model of walking where the process is described using sensory information and motor activities as state variables. Sensory signals are used as inputs to the ML. Since the muscles are operating as low-pass filters with respect to neural inputs, the prediction of the control signals driving muscles and the prediction of sensory signals are used as outputs from the ML. The inputs and outputs for the learning used for this study are obtained from simulation of a fully customized musculo-skeletal model described in details elsewhere [28]. The supervised learning task for an ML is to extract all invariant characteristics from the relationship between the provided inputs and outputs of the system (examples) and to store them in form of decision tree which can later produce approximated outputs when only inputs are provided. Since MLs generally do not have any limitations regarding the number of inputs and outputs, this approach is appropriate for multi-input-multi-output systems. The results provide good basis for design of robust control systems for FES - assisted walking.
Key words: Machine learning techniques, rule based control, artificial neural networks, fuzzy logic, entropy.