Adaptive Control Learning Based on a Similarity Measure

This work proposes a learning methodology to be employed in an adaptive controller strategy. It is based on a similarity approach along with a pole placement technique, by combining ideas from adaptive control and machine learning areas. The learning scheme is propped on the hypothesis that the current characterization of a given system can be achieved from the analysis of past similar behaviors. The main assumption is that data gathered from past experiments, during the operation, can be used on-to line reduce the uncertainty of a model that describes the system. As result, the strategy contributes to improve the performance of a controller based on that model. Two main steps are involved. In the first step a similarity analysis is performed, enabling to find in the historical a set of patterns (input/output time series) similar to the current condition. Then, in a second step, these time series are used to estimate the parameters of a linear model, that are employed afterwards in a pole placement control tuning. The applicability of the proposed approach is assessed on a benchmark nonlinear process, namely a continuous stirred tank reactor (CSTR), showing a better performance than with fixed PI and pole placement controllers.