Introduction Genetic programming is a domain-independent problem-solving approach in which computer programs are evolved to solve, or approximately solve, problems. Genetic programming is based on the Darwinian principle of reproduction and survival of the fittest and analogs of naturally occurring genetic operations such as crossover (sexual recombination) and mutation. John Holland's pioneering Adaptation in Natural and Artificial Systems (1975) described how an analog of the evolutionary process can be applied to solving mathematical problems and engineering optimization problems using what is now called the genetic algorithm (GA). The genetic algorithm attempts to find a good (or best) solution to the problem by genetically breeding a population of individuals over a series of generations. In the genetic algorithm, each individual in the population represents a candidate solut

Evolvable hardware (EHW) has attracted increasing attention since early 1990's with the advent of easily reconfigurable hardware such as field programmable gate arrays (FPGAs). It promises to provide an entirely new approach to complex electronic circuit design and new adaptive hardware. EHW has been demonstrated to be able to perform a wide range of tasks from pattern recognition to adaptive control. However, there are still many fundamental issues in EHW which remain open. This paper reviews the current status of EHW, discusses the promises and possible advantages of EHW, and indicates the challenges we must meet in order to develop practical and large-scale EHW. 1 Introduction Evolvable hardware (EHW) refers to hardware that can change its architecture and behaviour dynamically and autonomously by interacting with its environment. At present, almost all EHW uses an evolutionary algorithm (EA) as their main adaptive mechanism. One of the key motivations behind EHW is to learn from N...

Abstract. The RoboCup simulator competition is one of the most challenging international proving grounds for contemporary AI research. Exactly because of the high level of complexity and a lack of reliable strategic guidelines, the pervasive attitude has been that the problem can most successfully be attacked by human expertise, possibly assisted by some level of machine learning. This led, in RoboCup’97, to a field of simulator teams all of whose level and style of play were heavily influenced by the human designers of those teams. It is the thesis of our work that machine learning, if given the opportunity to design (learn) “everything ” about how the simulator team operates, can develop a competitive simulator team that solves the problem utilizing highly successful, if largely nonhuman, styles of play. To this end, Darwin United is a team of eleven players that have been evolved as a team of coordinated agents in the RoboCup simulator. Each agent is given a subset of the lowest level perceptual inputs and must learn to execute series of the most basic actions (turn, kick, dash) in order to participate as a member of the team. This paper presents our motivation, our approach, and the specific construction of our team that created itself from scratch. 1

We present a method of evolving analog electronic circuits using a linear representation and a simple unfolding technique. While this representation excludes a large number of circuit topologies, it is capable of constructing many of the useful topologies seen in hand-designed circuits. Our system allows circuit size, circuit topology, and device values to be evolved. Using a parallel genetic algorithm we present initial results of our system as applied to two analog filter design problems.

Techniques of evolutionary computation generally require significant computational resources to solve non-trivial problems of interest. Increases in computing power can be realized either by using a faster computer or by parallelizing the application. Techniques of evolutionary computation are especially amenable to parallelization. This paper describes how to build a 10-node Beowulf-style parallel computer system for $18,000 that delivers about a half petaflop (10 15 floating-point operations) per day on runs of genetic programming. Each of the 10 nodes of the system contains a 533 MHz Alpha processor and runs with the Linux operating system. This amount of computational power is sufficient to yield solutions (within a couple of days per problem) to 14 published problems where genetic programming has produced results that are competitive with human-produced results. 1.

High-level analog circuit design is a complex problem domain in which evolutionary search has recently produced encouraging results. However, little is known about how to best structure evolution for these tasks. The choices of circuit representation, fitness evaluation technique, and genetic operators clearly have a profound effect on the search process. In this paper, we examine fitness evaluation by comparing the effectiveness of four fitness schedules. Three fitness schedules are dynamic – the evaluation function changes over the course of the run, and one is static. Coevolutionary search is included, and we present a method of evaluating the problem population that is conducive to multiobjective optimization. Twenty-five runs of an analog amplifier design task using each fitness schedule are presented. The results indicate that solution quality is highest with static and coevolving fitness schedules as compared to the other two dynamic schedules. We discuss these results and offer two possible explanations for the observed behavior: retention of useful information, and alignment of problem difficulty with circuit proficiency. 1

. The basic properties of programmable analogue arrays are described and the problem of quantifying the fitness of an analogue circuit is discussed. A set of blocks appropriate for use in an evolutionary algorithm is described and results presented showing how an evolutionary algorithm using these blocks can learn to produce a given input-output characteristic. Finally an example is presented showing how the evolutionary algorithm can exploit any looseness in the specification of the desired characteristic. 1 Introduction The majority of work published so far on evolvable hardware has been carried out with digital logic systems. However the world is inherently analogue and the applications that many devices are put to involve analogue inputs and outputs. Some work has been done to exploit the inherently analogue nature of `digital' circuits, using properties of the circuit that were not envisaged by the designer. Whilst this enables devices to be used for tasks well outside th...

We examine the effectiveness of gradient search optimization of numeric leaf values for Genetic Programming. Genetic search for tree-like programs at the population level is complemented by the optimization of terminal values at the individual level. Local adaptation of individuals is made easier by algorithmic differentiation. We show how conventional random constants are tuned by gradient descent with minimal overhead. Several experiments with symbolic regression problems are performed to demonstrate the approach's effectiveness. Effects of local learning are clearly manifest in both improved approximation accuracy and selection changes when periods of local and global search are interleaved. Special attention is paid to the low overhead of the local gradient descent. Finally, the inductive bias of local learning is quantified.

Digital sound synthesizers, ubiquitous today in sound cards, software and dedicated hardware, use algorithms (Sound Synthesis Techniques, SSTs) capable of generating sounds similar to those of acoustic instruments and even totally novel sounds. The design of SSTs is a very hard problem. It is usually assumed that it requires human ingenuity to design an algorithm suitable for synthesizing a sound with certain characteristics. Many of the SSTs commonly used are the fruit of experimentation and a long refinement processes. A SST is determined by its “functional form ” and “internal parameters”. Design of SSTs is usually done by selecting a fixed functional form from a handful of commonly used SSTs, and performing a parameter estimation technique to find a set of internal parameters that will best emulate the target sound. A new approach for automating the design of SSTs is proposed. It uses a set of examples of the desired behavior of the SST in the form of “inputs + target sound”. The approach is capable of suggesting novel functional forms and their internal parameters, suited to follow closely the given examples.

A mathematical model for interactive sound synthesis based on the application of Genetic Algorithms (GA) is presented. The Evolutionary Sound Synthesis Method (ESSynth) generates sequences of waveform variants by the application of genetic operators on an initial population of waveforms. We describe how the waveforms can be treated as genetic code, the fitness evaluation methodology and how genetic operations such as crossover and mutation are used to produce generations of waveforms. Finally, we discuss the results evaluating the generated sounds.