Creating and Exploring Huge Parameter Spaces: Interactive Evolution as a Tool for Sound GenerationSkip other details (including permanent urls, DOI, citation information)
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 License. Please contact email@example.com to use this work in a way not covered by the license. :
For more information, read Michigan Publishing's access and usage policy.
Page 00000001 Creating and Exploring Huge Parameter Spaces: Interactive Evolution as a Tool for Sound Generation Palle Dahlstedt Innovative Design, Chalmers University of Technology SE-412 96 Goteborg, Sweden palle @design.chalmers.se Abstract In this paper, a program is presented that applies interactive evolution to sound generation, i.e., preferred individuals are repeatedly selected from a population of genetically bred sound objects, created with various synthesis and pattern generation algorithms. This simplifies aural exploration of huge synthesis parameter spaces, and presents a possibility for the sound artist to create new sound engines customized for this kind of creation and exploration - sound engines too complex to control in any other way. Different sound engines are presented, together with a discussion of compositional applications. It is also shown how this technique can be used to simplify sound design in standard hardware synthesizers, a task normally avoided by most musicians, due to the required amount of technical understanding. Keywords: interactive evolution, genetic algorithms, music composition, sound design, sound synthesis 1 Introduction Sound synthesis tools used by today's composers typically involve a large number of parameters, which makes them difficult to explore by hand. Synthesizers and software synthesis tools have great potential but are difficult to use. One way of exploring these huge parameter spaces, of sometimes hundreds of parameters, is by interactive evolution. In this paper I present a program, called MutaSynth, that applies interactive evolution to sound and pattern generation. It does not have any synthesis capabilities, but works by controlling an external sound engine by MIDI remote control. In this way it can be used with almost any synthesis tool. The external sound engine is controlled through a genetic representation of its parameter set and variations on these genomes are generated using a number of genetical operators. The resulting sound objects are auditioned and selected according to the user's aesthetical preferences. The process is repeated until a satisfying result has been reached. Since any synthesis tool can be used, a wide range of different sound objects can be evolved in a simple interactive process, such as techno loops, aperiodic electroacoustic sound material and traditional instrument sounds. The ability to control such huge number of parameters simultaneously opens up new possibilities of creating synthesis tools of a high complexity that would otherwise be very difficult to control. Two examples of such sound engines are presented, and also an application of interactive evolution on a standard hardware synthesizer, which greatly simplifies the sound design, even without knowledge about the underlying synthesis technique. There are many previous examples of experiments with genetic algorithms in music generation, most of them dealing with sequence generation. A good overview is given in (Burton and Vladimirova 1999). An example of an application in a musical work is (Jacob 1995). Maybe the most interesting is the work of Johnson (Johnson 1999), where parameter sets for a granular synthesis engine are evolved. This project is one in a series of experiments on the application of evolutionary algorithms to musical composition, including interactive evolution of score material and coevolution of sonic communication (see, e.g., (Dahlstedt and Nordahl 2001)). 1.1 Genetic Algorithms and Interactive Evolution During the last few decades, the simulation of biological processes has become a research field of growing importance. The area was named "artificial life" in the middle of the 1980's by Langton (Langton 1989). One of the main tools in this field is the genetic algorithm (Holland 1975), where artificial objects undergo selection and evolution. For this to work, the properties of objects such as images, sounds, or sorting algorithms, must be mapped onto a genome, which often consists of a string of numbers. Typically, the algorithm works like this:
Page 00000002 1. A random population is created. 2. The individuals of the population are evaluated according to some fitness criterion. 3. Those with highest fitness are replicated, often with some random modification (mutations), and/or mated with each other, to create a new generation of individuals. The old population is deleted. 4. The process is repeated from step 2 until a certain fitness threshold has been reached or until the increase in fitness stops. When evolving esthetically pleasing objects like sounds or images, it is difficult to automate the fitness criteria, since they cannot be explicitly defined. One solution is to let a human evaluate the fitness by some kind of inspection of the whole population before the selection and replication is made. This process is called interactive evolution, and was first considered by Dawkins (Dawkins 1986), who wrote a program generating simple drawings (biomorphs) that could be evolved to look like plants and insects. The purpose was to show how simulated evolutionary processes easily could give rise to complex forms reminiscent of actual biological form. The idea was taken up by Sims (Sims 1991), applying it to the evolution of two-dimensional graphics, where the objects were represented by hierarchical genomes consisting of mathematical formulae. For a number of other applications, see the book by Bentley (1999). A multi-dimensional parameter space, such as the parameter set of a synthesis algorithm, is an ideal target for interactive evolution. The individuals can be parameter sets for sound engines or pattern generators, like a synthesizer preset consisting of 100 parameters or more. You can navigate the huge parameter space without any knowledge of its underlying structure, following your esthetical preferences, and still have a very high degree of control. There are two creative processes involved. First you design a sound engine by hand, which is then explored by way of interactive evolution. Exploring a huge parameter space is essentially an act of creation, when the space is truly unknown. This exploration is carried out in an interactive process. Depending on the sound engine used, the result may be heard in real time. The exploration is not continuous but step-wise. Departing from a number of random sounds, you generate a number of variations on the ones you like. This process is repeated any number of times. In this way, you can search the parameter space for interesting regions in parameter space (i.e. interesting sounds), move between them instantly or in a desired time by interpolation. By combining different sounds, you can find other potentially interesting regions between them. 2. Implementation Working with MutaSynth involves two different modules - the sound engine which generates the actual sound, and the program, which manages the selection process and calculates the genetical operators. This section will concentrate on the program, how it works and how the sounds are represented genetically. Any possible sound generated by a certain sound engine can, using an engine-specific mapping table, be represented as a string of real numbers corresponding to the actual parameter values, ordered so that parameters related to each other (e.g. all filter parameters or all parameters controlling a certain oscillator) are located close to each other. We call such a string a genome, and every single real number in the string is called a gene. To make the parameters behave in a musical way, i.e. to maximize the percentage of good sounds generated, the genes are mapped to their corresponding synthesis parameters according to different translation curves, to make the most musically useful values more probable, while maintaining the possibility of more extreme values. An example is a vibrato amount, i.e., a low frequency modulation of the pitch of an oscillator. Too much modulation is often undesirable, and only values near zero are relevant for most applications. So, the gene range is mapped to the parameter range in a non-linear way to make very small values more probable. This technique is implemented using interpolated lookup tables. Any curve could be used, but simple exponential, logarithmic and cubic curves have proven most useful. To audition a sound, the corresponding genome is sent to the sound engine according to the mapping table and the translation tables, and the engine produces the sound. In MutaSynth, the working space is a population of nine sounds or genomes. The number nine is chosen because it corresponds to the number keys on the numerical keyboard, which is a convenient interface for sound selection, and because the auditioning process takes some time compared, e.g., to browsing images. In these contexts, a larger population is often used. If no suitable offspring is found, one always has the choice of repeating the last genetic operation with the same parents, to obtain more alternatives. As a starting point for the operations, the user can begin either with a set of randomly generated genomes or any previously stored genome. Sometimes it is also possible to import genomes, if the sound engine is capable of transmitting its parameter values. This may be useful for example when mating factory sounds in a synthesizer. 2.1 The Genetic Operators Genetic operators take one or two parent genomes and generate variations or combinations of them, resulting in a new population of genomes to be evaluated. In MutaSynth, the genetic operators used are mutation, mating, insemination and morphing. These are described individually below. Mutation and mating (also known as crossover) are standard genetic operations modeled after the genetic replication processes in nature. Insemination is a
Page 00000003 variation on crossover where the amount of genes that are inherited from each parent can be controlled. What I call morphing is a linear interpolation on the gene level. Every operator creates a set of new genomes that can be auditioned and further bred upon in the interactive process. Any sound can be stored at any stage in a gene bank, and the stored genomes can be brought back into the breeding process anytime, or saved to disk for later use. The parents used in an operation can be selected from several sources: a previously stored genome, either of the most recently used parents, any uploadable sound in the current sound engine or an individual from the current population (i.e., the outcome of the last breeding operation). A genome is really just a string of numbers, of constant length. Another sound engine would interpret these numbers differently. This means that a genome is meaningless without the sound engine it was created for, and it will not work with any other engine. It is sometimes useful to be able to prevent a number of genes from being affected by the genetic operations. For instance, when certain parameters of a sound (e.g., the filter settings) is good enough and the user does not want to run the risk of messing them up in further breeding operations, she can disable them, and they will stay as they are. If a gene is disabled, it will be copied straight from the first parent. Mutation. A new genome is generated from one parent sound's genome by randomly altering some genes. A mutation probability setting controls the probability of a gene to be altered and a mutation range sets the maximum for the random change of a gene. Together, these two allow control of the degree of change, from small mutations on every parameter to few but big mutations. Mating (Crossover). Segments of the two parent genomes are combined to form a new genome. The offspring's genes are copied, one gene at the time, from one of the parent genomes. A crossover probability setting controls the probability at each step to switch source parent. The starting parent for the copying process is selected randomly. Each parent will provide half of the offspring's genes, on average. The genes keep their position within the genome during this copying. Insemination (Asymmetrical Crossover). For a new offspring genome (Q), the following process is applied, based on two parent genomes (PI and P2): Pi is duplicated to Q, then a number of genes are overwritten with the corresponding genes in P2. An insemination amount controls how much of P2 should be inseminated in PI, and the insemination spread setting controls how much the genes to be inseminated should be spread in the genome - should they be scattered randomly or appear in one continuous sequence. If the insemination amount is small, the resulting sounds will be close in character to the sound of PI, with some properties inherited from P2. Morphing. A linear interpolation is performed on every gene of the two parent genomes, forming a new genome on a random position on the straight line in parameter space between the first parent (PI) and the second parent (P2). Manual Mutation. Manual mutation is not a genetical operator, but still something that affects the current genome. When the user changes a parameter on the synthesizer, the program is informed about the change and applies the change to the corresponding gene in the currently selected genome. The manual change then lives on through further breeding. Optionally, the manually changed gene can be automatically frozen, since a manual change is a strong decision. Manual mutations allow for the same level of direct control that the advanced synthesizer programmer is used to, and makes the method useful to both beginners and experienced users. Manual mutation may not be possible with all sound engines, depending on if they transmit parameter changes via MIDI. 2.2 User Interface MutaSynth is made to be simple. It is also designed to give quick responses to user actions, to minimize all obstacles in the creative process. Currently, the user interface looks like this: The display shows a number of boxes representing the population, the last used parents and the currently selected genome in the gene bank. The layout is chosen to correspond to the nine number keys on the computer keyboard. To listen to any individual from the current population, the user presses the corresponding number key, and the parameter interpretation of the genome is sent to the sound engine. The keys +, -, * and / invoke the different breeding operators. With these keyboard shortcuts the
Page 00000004 composer can keep one hand on her instrument and one on the number keyboard, for quick access to the operators and the individual sounds. Visual Representation of Genomes. The genomes are represented graphically, to aid the aural memory and to visualize the proximity in parameter space between genomes. For this purpose, I have developed a simple mapping technique. The gene values are interpreted as distances and angles alternatively (scaled to a reasonable range) and drawn as turtle graphics. The resulting worm-like line is scaled to fit in a picture box. Fig. 2: Visualization of three closely related genomes. When the different genomes are closely related, this is clearly visible (see fig. 2). Also, very often the graphics may have a resemblance to something figurative, which may aid your memory. Communication and Customization. All communication with MutaSynth and the sound engine, be it a hardware synth or another software, is by way of MIDI, using Midi Continuous Controllers (CCs) or System Exclusive Messages (Sysex). These messages can be customized to control virtually any hardware or software sound engine that supports MIDI remote control. This is done by creating a profile for that specific sound engine. A profile contains information on how to communicate with the synth, such as parameter-to-gene mapping, parameter ranges and communication strings. It is sometimes useful to have several profiles for the same sound engine, biased towards different subspaces of the parameter space or, musically speaking, towards different sound types. This is easily done by enabling different parameters subsets and specifying different parameter ranges, mappings and default values. 3. Sound Engines In this section I will describe three different sound engines that I have used with MutaSynth, two custom designed and one standard synthesizer, and discuss what kinds of sounds they can generate and how they can be explored with interactive evolution. The examples are chosen to show the wide range of possible applications of this technique. When designing sound engines for MutaSynth, there are two ways to go. Either you take an existing sound engine of some sort, such as a stand-alone synthesizer, a softsynth or a granular software engine. Then you try to make a gene-toparameter mapping that is relevant for the type of sound you want to produce, by carefully selecting parameters to be mapped and reasonable parameter ranges. Or, you build a sound engine from scratch, with the potential to generate the sounds you want, and probably many other sounds. This second approach is more flexible and open-ended, while the former is more time-efficient and can breath life into old awkward equipment. These examples are implemented in different environments, to show that MutaSynth can be used with any universal sound generation and processing tool, as long as it supports remote MIDI control of its parameters via Continuous Controllers or System Exclusive Messages. The first two examples may look confusing and unpredictable. This is because they are designed for another working method than the usual manual programming, allowing for all kinds of modulations to happen, maximizing the potential of the engine. Not all modulations will be active or influential all the time, depending on the genome. It would be very difficult to make any sound design by hand with these engines, but with interactive evolution it is quite possible. Sound examples of these and other sound engines can be found on http://www.design.chalners.se/palle/mutasynth. Fig. 3: The sound engine ToneMatrixC, as a screenshot from the Nord Modular Editor. Parameters included in the genome are marked with gray dots.
Page 00000005 3.1 ToneMatrixC This engine is implemented on the Nord Modular synthesizer (Clavia 1997), which is a stand-alone virtual modular synthesizer - basically a number of DSPs programmed by way of an external graphical editor. One can freely create and connect modules from a library of about 110 different types. Any parameter on any module can be assigned a MIDI Continuous Controller, which makes it easy to map up to about 120 genes to relevant parameters. ToneMatrixC is basically six sound generators (one in each column in Fig. 3) connected in a modulation matrix, so that each generator is modulated in different ways by a weighted sum of the others (the upper row of mixers). Each generator consists of an oscillator capable of four basic waveforms (sine, triangle, saw and square), a filter and some auxiliary modules. The modulation source is used, either as a pure audio signal or as an extracted amplitude curve, for different things in different generators such as ring modulation, frequency modulation and filter cut-off modulation. Each generator is played by an event sequencer track generating rectangular amplitude envelopes in a loop of 8 or 12 steps. The output is a mix of all six parts. Waveforms, pitches, the 30 modulation weights, filter parameters and almost all event sequencer steps are mapped to genes. All in all, 120 parameters are included in the genome, each spanning the full parameter range. This sound engine has potential for a wide range of rhythmic textures, from tonal blips to very complex feedback timbres, from snare drum march patterns to horror effects. Due to the complexity of the modulations, a large portion of the possible sounds are unusable because they are too harsh or noisy, but with careful breeding very interesting sounds can be found. Mating of different sounds works well, and breeding is easy, since the difference in sound is proportional to the mutation probability - small mutations yield small changes in the sound, and vice versa. and processing environment with flexible sampling capabilities. SampleMassage consists of four identical sample-players, each including an 8-step low frequency ramp oscillator controlling a virtual playback head in a sample look-up table. The sound material for the look-up table can be imported from any sound file, and the four players can have different sound material of different lengths. The ramp oscillators are synchronized and have a periodicity of about one to ten seconds. The ramp oscillator makes linear interpolation between eight equally spaced break points. Its periodic movement controls the playback of the sound in a very flexible way, creating interesting rhythmic and timberal patterns, and possibly pitch gestures, depending on the sound material used. From each sample-player, the eight ramp levels, the lowpass smoothing of the ramp wave, output amplitude and the start and end points of the used range of the sound are all included in the genome. Together with the global tempo, it amounts to 53 parameters included in the genome. They are grouped in the following overlapping groups: sample-player 1/2/3/4, ramp levels, start and end points, ramp smoothing and amplitudes. The nature of this sound engine is once again quite userunfriendly to a manual programmer, considering its great number of seemingly similar parameters, but with a little random search and breeding, wonderful grooves can be found. The genetical representation works well, since crossover operations really create audibly related combination sounds from very different parents and mutations work in a predictable manner. I have used this it with widely different sound material, from drum loops to voice recordings. Short sounds work best, or the playback head will sweep to quickly through the sound. Typically, I load a drum pattern into one sampleplayer, a bass pattern into another, and different tonal patterns into the others. The same sound can be loaded into all four, panned to different positions in the stereo field, creating very homogenous but complex textures. The character of the sound is very organic, since it is based on sweeps back and forth through the sounds, with no jumps or cuts. This sound engine also exists in an extended version, where each sample-player contains a large number of stored sounds. The genome contains an extra gene, selecting which sound to use. With this solution, the sonic potential of the engine is greatly enhanced. When the sound engines depend on stored sounds as raw material, a stored genome must be used with the same sounds when it is recalled, or it will have no meaning. A sweep through a different sound will obviously not sound the same, even if the break points are at the very same locations in the file. However, a genome may be used with different sounds if they are of the same lengths and have the same temporal structure. 3.2 SampleMassage Fig. 4: A schematic diagram of the SampleMassage sound engine. The SampleMassage engine is implemented in Native Instrument's Reaktor, which is a universal sound generation
Page 00000006 3.3 TX81Z The Yamaha TX81Z is a hardware synthesizer from the late 1980s, based on 4-operator FM synthesis. It is mentioned here to show that standard synthesizers are perfectly usable as sound engines with MutaSynth, and that MutaSynth is very suitable as an interactive synth editor, requiring little or no knowledge about the meaning of individual parameters. The genome consists of the TX81Z's major voice parameters, grouped in the following partly overlapping categories: globals, envelopes, operator frequency ratios, operator waveforms, operator levels, LFOs, patch name, opl/op2/op3/op4 parameters, keyboard scaling and modulation sensitivities. All in all, there are 110 genes. The parameter values are communicated to the synth by way of MIDI System Exclusive Messages. Starting from factory sounds, interesting variations can be found in a few generations. Best results are achieved when a few groups are evolved at a time, e.g. the operator ratios but not the operator levels, or vice versa. Evolving one operator at a time, leaving the finished ones enabled is also a fruitful strategy. Random exploration of all parameters does not result in many useful sounds, though, due to the small fraction of good-sounding frequency ratios in FM synthesis. Crossover of sounds proved to be very fruitful, with clearly audible characteristics inherited from both parents. 3.4 Observations There are different ways of thinking when constructing sound engines to use with interactive evolution, depending on its purpose. Firstly, I need to know what kind of sounds I want to produce. Not an exact description, but I need to know how they should be generated, what techniques to use, what kind of parameters should be mapped to genes and such things. In general, I have to think about the potential of the engine, considering the consequences of the whole range of each parameter and the probabilities of related parameters to have meaningful combinations of values. These are crucial choices - there is nothing like a universal synthesis engine, except maybe to interpret a huge genome as a wave file, but that would be quite useless because of the infinitely small portion of the possible genomes that will actually make any meaningful sound. Most will be just noise. The more universal I make the sound engine, the smaller the portion of useful sounds. If I try to cover many different types of sounds, I will also allow the generation of strange sounds in between the useful areas in parameter space. These may be interesting, but probably only a very small minority. The genetical representation of the sound, i.e., the mapping of the genes to the sound generation parameters, is crucial and careful attention has to be paid to the choices of parameters and ranges. Sometimes parameters are dependent on each other, like in FM synthesis. Then the modulator frequency should be an integer multiple or fraction of the carrier frequency for a harmonic result. This is of course a simplification, to illustrate the problem. There are several ways to avoid this. One is to design the sound engine so that one gene maps to the carrier frequency and another to the integer ratio between them. Then the probability for useful sounds increases, but the results are also more limited. Another way is to start the evolution from genomes that sound good, then limit the use of big mutations, concentrating on mating with other good genomes, and possibly use some insemination. Then the ratio between the dependent genes' values will be inherited from one of the parents, except if a crossover occurs between them, which is unlikely if they are located close to each other in the genome. The disadvantage of this solution is that it puts restrictions on the working method and requires the user to be aware of what is happening at the genetic level, which may not always be the case. Yet another way of dealing with very large sound engines is to divide the parameters into relevant groups, applying interactive evolution to one group at a time. If the sound engine is an FM synth, one could first evolve the frequency ratios, then the envelopes and modulation parameters, or freeze a set of predefined, plausible FM ratios, while evolving the other parameters, such as modulation depths and envelopes (as in the TX81Z example). The grouping technique is also effective in limiting the possible damages of bad mutations, by protecting the finished parts - according to your fitness evaluation - of the genome. Grouping may not always be possible, though. For example, if the sound engine is a homomorphic network of intermodulating oscillators, it may be very difficult or impossible to evolve e few parameters at a time, since they all depend so much on each other. Many synthesis tools keep the event level and the timberal level separate. Personally, I like to use sound engines that combine sound synthesis and pattern generation in one way or another, instead of just evolving keyboard playable sounds and then playing them. It is possible to parameterize both structural patterns on the macro level and the timbre at the micro level, and evolve them simultaneously. In this way, you can explore a continuum in both these dimensions at the same time, which allows for more complex audible relationships between sound objects. 4. Applications Considering the wide range of possible sound engines and synthesis techniques, there are a number of applications for this program in the field of sound design and electronic composition and performance. 4.1 Material Generation MutaSynth can be used on the sound level, for developing sound objects, loops, timbres and structures to
Page 00000007 be used in manual compositional work. In this case, it offers a way of investigating the possibilities of a certain sound engine. A musical advantage is that the material created during a breeding session often is audibly interrelated to quite a high degree, which opens up for compositions based on new kinds of structural relationships. 4.2 Interactive Evolution of Synthesizer Sounds MutaSynth can be used for programming almost any MIDI synth, without any knowledge about the underlying synthesis techniques. This depends on the existence of a suitable profile for the synth model in question, defining a genome with a relevant mapping to the parameter set in the synth. These profiles are quite easy to create, given the parameter set and communications protocol of the synthesizer in question. The genetic operators provide a simple way of trying out combinations of existing sounds, finding variations or just exploring the parameter space of the synth. Doing this on normal off-the-shelf synths I have found sounds that I never thought were possible. Many computer based synth editors offer a randomization feature, sometimes with a random step size, corresponding to genetic mutations. They do not, however, provide anything like crossover or morphing, to explore the parameter space between existing sounds, or an interface for iterated interactive selection of the preferred sounds. 4.3 Live Evolution With a suitable set of sound engines synchronized to a common clock, several layers of musical material can be evolved simultaneously, live or in the studio. Either by evolving one layer while the others play, possibly with headphone monitoring, and sending it out to the speakers when a nice pattern has been found, or by evolving over many sound engines at the same time, like a giant genome consisting of all available knobs/parameters. Also, since the gene banks of MutaSynth are controllable by MIDI Program Change messages, the playback of previously evolved stored sounds is quite easy to integrate in a live set-up. I think this would be a very powerful tool for DJ:s and composers of music for the dance floor. 4.4 Installation Version Interactive evolution of musical textures may also be considered interesting in its own right, and an installation version has been developed (Dahlstedt 2001), which lets the user evolve continuous rhythmic patterns of different styles in three layers with a simplified interface, i.e. with mutation rate, mutation probability, morphing and a few other features removed. This installation has really proved the simplicity of the interface, since musically untrained users easily evolved complex techno patterns, adjusting 360 parameters to suit their taste without even knowing it. The installation version also implements automated gradual crossover of stored patterns in stand-by mode, when nobody is interacting with it, resulting in continuously new textures. 5. Discussion Interactive evolution as a compositional tool makes it possible to create surprisingly complex sounds and structures in a very quick and simple way, while keeping a feeling of control. I am still the composer, rather than a slave of the application. The process is interactive and gives immediate aural feedback, which is crucial for creative work, while it allows using very complex sound engines that would be difficult to use productively in any other way. Interactive evolution does not limit the creativity by imposing templates or modes on the result, like many other creativity aids such as document templates, groove templates, chord sequences or phrase libraries. The sound engine defines a huge space, which is searched without restrictions in all dimensions with the help of chance and esthetic judgment. MutaSynth can be combined with manual control, through manual mutation and freezing of groups of genes. An expert can easily make a parameter change in the middle of the breeding process, and a manually designed sound can be good starting points for further breeding. I like to be surprised by chance and emergence, while other composers may prefer total control. They may feel uneasy working with a technique like this, since it is not a tool for achieving a predefined goal, but rather searching the unknown, with your ears wide open. So this is maybe not a tool for those who know exactly what they want. In my view, though, creativity involves a certain amount of unpredictability and surprise. There seems to be an upper limit of the duration of the sonic fragments that can be created with MutaSynth, without loosing their musical meaning. My experience is that it is possible to evolve either continuous sounds that form one small part of a dense structure, or short selfcontained structures that can sustain interest by themselves for a short time, but not longer sections or whole pieces. A reasonable duration to expect from these evolved structures could be up to maybe fifteen seconds. To search an unknown space differs from a typical optimization process guided by an explicit fitness criterion. It is more about finding local maxima in an implicit fitness function, i.e. to find pleasing sounds, than to locate a global maximum. Ending up with a good sound one did not know one needed might be a plausible result. If the goal is to design a good oboe sound and a good trumpet sound is found on the way, it is just a bonus. I often save sounds during the whole breeding session, the last not being any better than the first, but different. Together they form an
Page 00000008 interesting progression, often appearing in that order in the finished composition. The way is the goal. In my own work, the impact of using MutaSynth has been significant. I often use it by evolving lots of sounds, sometimes in several synchronized layers, storing all the interesting genomes in the gene bank. Then I record the sounds and edit them together to make a composition from it. Because of my personal belief and interest in the unexpected, this is a very suitable tool. The act of interactivity is blurring the border between improvisation and composition - it is more a controlled way of improvisation, with inherent selection of the best takes. Also, the continuous audition of new material is changing the focus from concept to result, relying heavily on the judgment of the composer. Composition is turned into a selection process, at least in the material-creation phase. Still, of course, the sound engines have to be created by hand, which is a highly creative and stimulating task, very similar to composing. Many crucial decisions are actually taken in this early phase, since the sound engine so definitely sets the boundaries of the sound world to be used. 6. Conclusion and Future Work It is clear that interactive evolution of parameter sets is a powerful tool for musical creation, provided adequate genetical mapping and a powerful sound engine. The technique could be even more powerful if evolution of the sound engines themselves was possible, requiring a more complex genetical data structure. One way could be to evolve wave-generating mathematical expressions, stored in tree structures. Another could be to implement a data structure representing the sound generation and processing modules, parameters and possible connections of a generalpurpose sound generation tool, including a database of relevant connections (to exclude the large amount of nonsounding module combinations). Both these implementations would allow evolution of arbitrarily complex sound objects. Waveform data could also be included in such a hierarchical genome, and possibly also be subject of genetical operators, creating new waveforms and timbres. Another related technique is to evolve scores or MIDI files, using tree structured event-based expressions. One implementation of this is currently under development, and has shown great potential so far. Automation of fitness criteria would be an interesting extension, allowing for batch evolution of sound objects. This could be based either on some kind of description of the desired sound, or by example, making the program evolve sound objects as close as possible to a given sound file, which if successful would make available a parameterized version of a recorded sound, which would open up new ways of editing sound. References Bentley, P. J. (ed.). 1999. Evolutionary design by computers. Morgan Kauffman. Burton, A.R. and Vladimirova, T. 1999. Generation of Musical Sequences with Genetic Techniques. Computer Music Journal 23:4 (1999), 59-73. Clavia DMI AB 1997. Nord Modular (a virtual modular synthesizer). Stockholm, Sweden. http://www.clavia.se. Dahlstedt, P. and Nordahl, M. 2001. Living Melodies: Coevolution of sonic communication. Leonardo Journal 34:3 243-248, 2001. Dahlstedt, P. 2001. A MutaSynth in Parameter Space: Interactive Composition Through Evolution. To appear in Organised Sound 6(2). Dawkins, R. 1986. The Blind Watchmaker, Essex: Longman Scientific and Technical. Holland, J. H. 1975. Adaptation in natural and artificial systems, Ann Arbour, MI: The University of Michigan Press. Jacob, B.L. 1995. Composing with Genetic Algorithms in Proceedings of CMC 1995. San Francisco, CA: International Computer Music Association. Johnson, C.B. 1999. Exploring the sound-space of synthesis algorithms using interactive genetic algorithms, in A. Patrizio, G.A.Wiggins and H.Pain (eds.) Proceedings of the AISB'99 Symposium on Artificial Intelligence and Musical Creativity. Brighton: Society for the Study of Artificial Intelligence and Simulation of Behaviour. Langton, C.G. (ed.). 1989. ALIFE I, Proceedings of the first international workshop of the synthesis and simulation of living systems. Addison Wesley. Sims, K. 1991. Artificial Evolution for Computer Graphics. Computer Graphics 25, 319-328.