Nature-Inspired Computing for Control Systems

This chapter presents a new biologically inspired technique for automatically compensating the light spot deviation from the normal position for laser spot trackers. The method is based on hardware implementation of the spiking neural networks which provides fast response due to real time operation and ability to learn unsupervised when they are stimulated by concurrent events. For increasing the biological plausibility of the method, the spiking neural network controls the contraction of shape memory alloy (SMA) actuator wires that operates as the muscular fibres. These SMA wires are the most suitable actuators for being controlled by the electronic spiking neurons because the contraction force increases naturally with the spiking frequency. From our knowledge the laser spot tracking using spiking neural networks was not performed previously. Moreover, other original ideas represent the use of analogue implementation of the spiking neural networks for real time operation as well as the SMA actuator wires for more biological plausibility. To validate this method we implemented in hardware a spiking neural network structure that processes the input from a one dimensional photodiode array and controls a positioning system based on SMA actuator wires. The results show that the spiking neural network is able to detect the one-dimensional spot motion and to adapt the response time by Hebbian learning mechanisms to the spot wandering amplitude. Moreover, by driving two antagonistic SMA actuator wires the system is able to track the laser spot with low response time and acceptable precision. These results are encouraging to develop bio-inspired low power spot tracking system for enhancing the receiving accuracy in free space optical communications or for enhancing the efficacy of the photovoltaic systems. Moreover, the light tracking principle based on spiking neural networks and SMA wires can be successfully used in implementation of the light tracking mechanism of an artificial eye.

The main purpose of introducing ambient assistive living (AAL) robots is to assist the disabled and elderly people at home. In recent years, this field has evolved quickly because of the enormous increase in computing power and availability of the improved variety of sensors and actuators. However, design of AAL robots control system is a huge challenge, which require solving issues related to two classes: design of mechanical structure and development of an efficient control system. In this chapter, we focus on the latter topic, since even relatively low quality hardware can be used for solving sophisticated tasks if the software control it correctly. The chapter starts by giving a vision of what heterogeneous AAL robots is supposed to look like and how a human is to act, navigate and function in it. Particularly, we investigate the effect of artificial neural network (ANN) based control techniques for AAL robots. To enhance the accuracy and convergence rate of ANN, a new method of neural network training is explored, i.e., grey wolf optimization (GWO). Moreover, we provide an overview of applying emerging metaheuristic approaches to various smart robot control scenarios which, from the author’s viewpoint, have a great influence on various AAL robot related activities, such as location identification, manipulation, communication, vision, learning, and docking capabilities. The findings of this work can provide a good source for someone who is interested in the research field of AAL robot control. Finally, we concludes with a discussion of some of the challenges that exist in the AAL robot control.

This chapter brings forth the practical aspect of using genetic algorithms (GAs) in aiding PID (Proportional-Integral-Derivative) control design for real world industrial processes. Plants such water tanks, heaters, fans and motors are usually hard to tune on-site, especially after prolonged use of the equipment when degradation of performances is inevitable, while plants like seismic dampers have inherent nonlinear behaviors that make formal controller design difficult at best. Therefore, this chapter introduces a series of practical steps that can be taken by control engineers in order to (re)design viable PID controllers for their plants. This chapter describes how genetic algorithms can be applied to problems in control systems and model identification. Considering the plant inputs and outputs that can be observed during functioning, we offer a quick method for identifying model parameters, which can be used later by the genetic algorithm to find a suitable controller. While, in formal control theory, a raw estimation of the model parameters can significantly reduce the performance of a real-world system, the genetic algorithm method can find suitable controllers quickly and efficiently, offering access even to performance criteria that is hard to quantify in classical design procedure, such as integral indexes. The applicability of the GAs in real world problems is outlined through case studies that take into account the particularities of each system, from first to second order responses, the absence or presence of time delay, nonlinearities, constraints and controller performance. The steps performed in the case studies show how GAs have made the jump from their origins to a practicing engineer’s toolbox (GAOT-ECM in this case, a Genetic Algorithm Optimization Toolbox Extension for Control and Modeling). Moreover, a comprehensive analysis is performed, that takes into account both the various performance criteria, and the tuning parameters of the genetic algorithm, over the obtained models and controllers. The influence of the GA parameters is discussed in order to help practitioners choose the best suited GA configuration for their particular problem. In all, this chapter offers a comprehensive step-by-step application of genetic algorithms in industrial setting, from plant modeling to controller design.

This chapter presents a Hybrid Elite Genetic Algorithm and Tabu Search (HEGATS) to design optimal fuzzy controllers for multi input multi output (MIMO) nonlinear system. The principle of the proposed algorithm is to seek the elitism by GA and introduce it in the TS algorithm as initial solution in order to find the best fuzzy rule base of the fuzzy controller. The fuzzy rule base of the fuzzy controller is tuned for optimal control performance using HEGATS by minimizing the mean square error. The proposed algorithm is tested for control of a helicopter model simulator and a double inverted pendulum. Simulation results proved the effectiveness of the proposed algorithm.

This work presents a recent contribution regarding the application of multi-layer perceptron neural networks (MLP-NNs) to the fitting of complex piecewise affine control and observation laws in constrained linear systems. Such input-output maps arise from the imposition of the optimal contraction rate trajectory for the system state, or error in the observer case, within a given invariant polyhedral set that enforces the systems constraints, or in the case of optimal control laws in model predictive control (MPC). Although an offline law can be computed via multi-parametric optimization in the state feedback case, the number of regions that define the piecewise affine law can be very large, resulting in high hardware storage requirements and difficulty in quickly locating the state. On the other hand, online laws are usually more expensive in the runtime sense and can become infeasible for application in fast dynamics systems, unless an advanced, expensive processor is employed. From the data obtained by the simulation of online laws, MLPs can be trained to emulate such maps; thus, they can replace online computation and drastically reduce the runtime. Two numerical examples, one of which is based on a two-tank hydraulic system model, are presented to illustrate the proposed approach, with detailed design for constrained error estimation as well as static and dynamic output feedback.

In this chapter, a novel adaptive neuro-fuzzy backstepping control scheme is developed for induction machines with unknown model, uncertain load-torque and nonlinear friction. Neuro-fuzzy systems are used to online approximate the uncertain nonlinearities and an adaptive backstepping technique is employed to systematically construct the control law. The proposed adaptive fuzzy controller guarantees the tracking error converge to a small neighborhood of the origin and the boundedness of all closed-loop signals. These neuro-fuzzy systems are adjusted on-line according to some adaptation laws deduced from the stability analysis in the sense of Lyapunov. Compared to previous works, the proposed controller can effectively deal with the induction motors drives with both unified nonlinear frictions and (structured and unstructured) uncertainties. In fact, this present work can be seen as a non- trivial extension of the previous works. Simulation results are provided to demonstrate the effectiveness of the proposed control approach.

Artificial intelligence has opened new alternatives to control non-linear systems. One of the most important methods is the fuzzy logic controller, which is constructed with linguistic rules; however, it is normally based on the knowledge from human experts, so the linguistic rules and membership functions are not optimized. In the case of Quadrotors, it is important to acquire the knowledge from human experts because they are able to naturally describe the controller for this non-liner system by linguistic rules in a complete form but the memory space and processing time have to be minimum in the hardware implementation. Hence, a neuro fuzzy controller (ANFIS) is used in order to reduce the number of linguistic rules and membership functions and to preserve the surface between inputs and outputs in the controller for Quadrotors. Hence, the real time controller improves its response. This proposal keeps the basic idea of getting the knowledge from human experts and then ANFIS can be implemented in the real time hardware in the Quadrotor. The results show that this methodology is an excellent option to control Quadrotors.

This chapter presents a new algorithm inspired in the human visual system to compute optical flow in real-time based on the Hermite Transform. This algorithm is applied in a vision-based control system for a mobile robot. Its performance is compared for different texture scenarios with the classical Horn and Schunck algorithm. The design of the nature-inspired controller is based on the agent-environment model and agent’s architecture. Moreover, a case study of a robotic system with the proposed real-time Hermite optical flow method was implemented for braking and steering when mobile obstacles are close to the robot. Experimental results showed the controller to be fast enough for real-time applications, be robust to different background textures and colors, and its performance does not depend on inner parameters of the robotic system.

Reinforcement learning methods can be computationally expensive. Their cost is prone to be higher when the cardinality of the state space representation becomes larger. This curse of dimensionality plays an important role on our work, since gait generation by using more degrees of freedom at each leg, implies a bigger state space after discretization, and look-up tables become impractical. Thus, appropriate function approximators are needed for such kind of tasks on robotics. This chapter shows the advantage of using reinforcement learning, specifically within the batch framework. A neuroevolution of augmenting topologies scheme is used as function approximator, a particular case of a topology and weight evolving artificial neural network which has proved to outperform a fixed-topology network for certain tasks. A comparison between function approximators within the batch reinforcement learning approach is tested on a simulated version of an hexapod robot designed and already built at our undergraduate and graduate students group.