Evolvable Hardware

One of the greatest challenges in Evolvable Hardware—in particular when dealing with intrinsic evolution—is to monitor and assess the performance of an (evolving) circuit or system in order to arrive at a solution that ultimately meets the design specifications. In order to be able to accurately assess performance it is necessary to accurately measure not only input/output characteristics, but also properties such as power consumption, resource consumption and temperature. From these measurements desired performance characteristics can then be calculated; the formula (or algorithm) that achieves this is called the fitness function. Once more borrowing terminology from biology, the fitness of a design under test represents its performance on one or more previously defined objectives. The fitness value is then used by the evolutionary algorithm to rank a population of candidate solutions (individuals), hence to decide which ones shall survive and produce offspring, and which ones shall be discarded. Note that although this chapter is mostly about electronic systems, many of the problems addressed are generic to all evolved physical systems and just the way certain properties manifest themselves changes. For example, in electronic circuits “device variability” occurs as a result of the manufacturing process and the equivalent in biological organisms would be “cell diversity”, which occurs due to how growth takes place.

Inspired by a seminal paper (Higuchi et al, 1993), many researchers have started to work on evolutionary design and optimisation of combinational digital circuits. In this method, electronic circuits encoded as strings of symbols are constructed and optimised by the EA in order to obtain a circuit implementation satisfying the specification. In order to evaluate a candidate circuit, a reconfigurable circuit (or its simulator if evolution is performed using a circuit simulator) is reconfigured using a new configuration created on the basis of the chromosome (i.e. configuration string) content. The configured device is then evaluated and its behaviour is compared with the desired behaviour. The fitness score is calculated, which reflects to what extent the candidate circuit satisfies the specification. The main reason why evolutionary circuit design has been studied and developed is its ability to (i) provide novel designs hardly reachable by means of conventional methods; (ii) deliver good solutions for problems where the specification is inherently incomplete and no golden solution exists; and (iii) achieve adaptation/fault tolerance directly at the hardware level. Recent overviews of applications and design methods can be found, for example, in (Sekanina et al, 2011; Haddow and Tyrrell, 2011; Sekanina, 2012).

Medicine is an important and challenging area for machine learning and computational intelligence approaches. Evolutionary algorithms, in particular, offer a number of advantages when solving problems in this domain. For instance, unlike decision trees and neural-network-based approaches, they are not tied to particular solution representations. This flexibility is important for problems where the aim is to model specific biological or pathological processes, and it also provides scope for using interpretable models, in which evolved solutions can be mined for information. Another important advantage is the ability of evolutionary algorithms to explore a relatively large space of solutions, without requiring a priori knowledge of the structure of a problem. This means that evolutionary algorithms can often be used as a discovery tool, providing new fundamental knowledge to clinicians. For these reasons and others, over the last decade there has been growing interest in the use of evolutionary algorithms in medicine (Smith and Cagnoni, 2011).