Application of Echo-State Networks to the Position Control of Shape-Memory Alloys

Abstract

Shape-memory alloys (SMA) have the ability to generate strain in response to temperature change. However, the relationship between applied temperature and crystalline phase is hysteretic. Obtaining an explicit representation of this highly nonlinear phenomenon for the purpose of control consumes time and effort. Also, the identification process is subject to uncertainties, and furthermore, the dynamic properties of SMAs may change during its lifetime which reduces the reliability of the identification. With this in mind, we employ an online learning control framework for the position control of an SMA wire. The online learning control framework performs an inverse learning of the plant based solely on the input and output signals, and uses this information to generate control inputs. Thus, an explicit representation of the plant is not required and as a result from online learning, the controller adapts to changes of the plant. The specific method with which the inverse learning is employed, is by the use of echo-state networks (ESN). ESNs are a class of recurrent neural networks and are distinguished by their large number of hidden nodes often referred to as a dynamic reservoir. While the originally proposed method of constructing this dynamic reservoir relies on a stochastic sampling process, recent studies have suggested that using a simple and deterministic reservoir also provides sufficient performance. Here, we also investigate the impact of using such simple and deterministic reservoir structure within the online learning control framework. Experiments of the online learning control framework conducted on an SMA wire are presented.