Cefn Hoile
Black Shoals : Evolving Organisms in a World of Financial Data

Abstract

In the Black Shoals project, artificial evolution was employed to design articulated creatures which interact with a world of real time financial data. Their phenotypes are composed of multiple interacting elements in a discrete time simulation of Newtonian physics. The couplings formed between evolved, recurrent neural nets and bodily components generate co-ordinated, efficient strategies for the exploitation of their environment. The ecosystem in which they are embedded implements a decentralised evolutionary algorithm, applying competitive selective pressure to the population through conservation of energy. It also couples them to a larger complex system. The opportunities in their world derive from real time stock market information fed from a live Reuters data-feed. The genotype to phenotype mapping employed offers the features of redundancy, neutrality and epistasis. Support is also provided for the ad hoc inclusion of new creature components without modifying the evolutionary algorithm. The focus of this paper is the design and implementation of the ecosystem-based population management, the genetic encoding and developmental process of the creatures.

THE BLACK SHOALS PROJECT

In nature, organisms exploit the complex dynamics of their multiple, interacting components to generate goal-oriented behaviour in a complex, changing world. The number of possible combinations and couplings of biochemical elements is astronomically huge. Despite the high dimensionality of this phenotype space, natural evolution is able to create highly complex coupled systems. A massive range of organisms has arisen, each with a different strategy for survival and reproduction. This innovation is achieved through an open-ended, decentralised evolutionary process – a result of limited resource availability combined with co-evolutionary interactions.

This report describes the design features of the evolutionary algorithm used in the Black Shoals arts project, intended to mirror some of the aspects of real world evolution described above. In this project, a diverse ecosystem was created, where each creature is a coupled system of simple components, operating in a resource constrained world.

Black Shoals was commissioned as a live display of both market and ecosystem dynamics as part of the “Art Now : Art and Money Online” exhibition, and ran at the Tate Gallery for three months. A key element of the artwork was to create a world of artificial 'speculators' able to 'feed' on the shifting real-time patterns of the world's stock markets.

In the display, the top 4000 traded public companies are each represented as individual stars projected onto a planetarium-style dome. The companies’ stars are clustered according to the correlation statistics of their stock histories. Starbursts are generated whenever trades for an individual company are received from the market, leading to cascades of illumination when live news events trigger trading activity in the various clusters.

The creatures occupy the ‘2d’ curved surface of the dome. The correlation process which clusters the stars generates meaningful neighbourhoods for the embodied creatures in the space to explore and seek food. Trades in the shares of a given company accumulate at its corresponding star, providing pockets of energy for the creatures to consume. The organisms’ survival depends upon navigating the surface of the dome in the most efficient way, competing with each other to acquire the limited energy which is fed into the system. This also provides the selection pressure for an evolutionary algorithm, since the metabolic, reproductive and kinetic costs incurred by the individuals in the population must be balanced by their energy intake in order to remain alive.

This paper focuses on the evolutionary dynamics, genetic encoding and developmental process through which lifelike creatures emerged unsupervised from a world of simple interacting components.

The ecosystem defines the challenge that the creatures face, encouraging purposeful behaviour to arise. It also gives rise to the self-regulated management of population.

The genetic encoding and developmental process give rise to features such as redundancy, neutrality and epistasis which could assist the creatures in developing complex behaviours.

Support is also provided for the ad hoc inclusion of new creature components without modifying the evolutionary algorithm.

THE CREATURES

The creatures are composed of point masses jointed together by light elastic rods, embedded in the surface of the dome. In turn, these physical elements link together with sensors, neurons and actuators. (See figure below and next section).

Their phenotype represents a dynamic system within a discrete-time simulation of Newtonian physics. The evolutionary process tunes the interconnections and parameters of the creatures’ physical elements, and those of the recurrent neural net, to produce efficient neuro-mechanical systems capable of goal-oriented behaviour.

Selective algorithm

The selective algorithm which drives the evolution of creature strategies is an approximation to the dynamics of the natural world. Initial creatures are randomly composed from basic elements. There are no generational cycles or strict periods of evaluation and reproduction; both are continuous, and a by-product of the energetic constraints of the physics imposed on the organisms. The success of a creature’s strategy depends upon acquiring sufficient energy from the environment through the foraging of food to compensate for the energy dissipated by the creature’s foraging and mating activities.

Energy Acquisition

Creatures may acquire energy by making contact with one of the 4000 ‘stars’ in their environment. Each star corresponds to a specific company on the world’s stock markets, and contains an energy value which reflects the live activity of that company’s stock within the world’s markets.

In order to survive and reproduce, a creature must find the zones of trading activity within the star-space before its competitors. This co-evolutionary dynamic provides the selective pressure, encouraging innovation and increasing the efficiency of gaits and sensorimotor control strategies found in the population over time.

The distribution of energy through the dome is a strict reflection of the comparative activity of the various companies. However, the net rate of energy input is normalised to prevent extremes of population explosion and collapse. When a trade is received from the live Reuters data feed, the energy accumulated at the corresponding star is augmented according to the monetary value of the shares traded. This value is normalised by a running average of trades received during the 1000 preceding simulation steps. Whilst extreme changes in trading activity still lead to short-term boom and bust periods of energy input, this damping effect helps to maintain a stable population.

When the world’s markets are closed, and hence trading has ceased, the same net input of energy is instead randomly distributed between the 4000 stars.

Energy Dissipation

There are several ways in which a creature’s activity leads to a dissipation of its acquired energy.

Metabolism

In nature, the maintenance of bodily structure in the face of increasing entropy demands a flow of energy. This aspect is reflected in the dynamics of the creatures’ world.

However, there are additional benefits for the control of the computational demands of the creature’s universe. The evaluation of each element incurs a computational cost on the processor which may lead to an unacceptable slowdown in the update speed of the discrete time simulation as creatures become more complex. In order to regulate the increasing complexity of the creatures, a continuous metabolic cost is due to each of the components which make up the body of the creature. Thus, it is in the interests of creatures to add components to their phenotype only if the component provides an energetic benefit. In the exhibit, this cost was tuned to achieve a target update speed of 10 frames per second to provide the illusion of continuous motion to the human eye.

Kinetic Energy

A kinetic cost is due to the muscular action of the creatures, calculated according to Newton’s laws [force x distance moved]. This constraint encourages the development of ‘lifelike’ gaits, selected according to their efficiency in traversing the dome. The damping friction inherent in the elastic rods of which the creatures physical bodies are composed may be thought of as dissipating this energy. Another means of losing kinetic energy is the fixing of one of the creatures’ point masses to the dome surface. When this foot-mass is fixed to the surface, all of its kinetic energy is lost, (effectively infinite friction).

Mating

A mating cost is due to reproductive acts. Creatures which have sufficient energy may choose to become mothers, using a special actuator, or reach a threshold of 4 times their birth energy. When they give birth, the mothers contribute the equivalent of their own birth energy to the creation of their offspring. The new individual is born within the neighbourhood of its mother, and is the result of a recombinant act with a father chosen at random from its own species (see later). This may be likened to the reproductive habits of creatures with sea-borne gametes, in which fertilisation is not an explicit act between co-located creatures. However, in Black Shoals, individuals have no specific gender, and may adopt either role – motherhood deliberately, fatherhood unknowingly.

Population Management

Whilst the principle of energy conservation provides an upper limit to the number of members in the population, and their complexity, there is still no guarantee that any of the individuals will be able to sustain their energy through foraging.

Initial populations are seeded randomly. As a result. However, since the competition is similarly inept, and hence food is plentiful,

Initial seed individuals, randomly initialised, have behaviours ranging from complete inertia to wild thrashing. They have badly designed bodies, and demonstrate no coordinated sensorimotor response. However, since a continuous flow of energy is introduced into the space, even the simplest foraging approaches will finally succeed in a world with no competition, which will contain large unexploited deposits of accumulated energy. Slight improvements in body plan and control over their neighbours can enable a creature to survive and reproduce. Over time as the population improves, the coevolutionary pressures demand more and more competence, and lead to the creation of more complex and well designed creatures.

An example of a simple foraging approach employs an articulated element attached to a permanently fixed point. The untuned muscular control of the armature causes it to swing wildly, clearing food from a circular area. The control to detach this foot comes later, followed by more complex, co-ordinated forms with multiple points of attachment, energy-efficient gaits and sensory responses.

Species and Sexual Isolation

The reproductive isolation of species allows a diversity of distinct strategies to co-exist, compete and evolve within the same ecosystem. It prevents the recombination of genomes with radically different configurations, which would producing large numbers of non-viable mutant offspring, and will finally lead either to extinction, or to the elimination of diversity.

At the genesis of this ecosystem, 50 individuals are introduced into the ecosystem with randomly initialised genomes. Each organism introduced (rather than born naturally) is the firstborn or seed of a new species; Species boundaries are maintained by enforcing the rule that only creatures who share a common firstborn ancestor may reproduce together. Since there is only one creature in each new species, the first reproductive act takes place asexually.

In order to maintain a minimum of 10 species, retaining the richness of the ecosystem display despite occasional extinctions, a circular queue of 50 saved species is maintained on disk. When the number of species drops below 10, saved species are reintroduced.

The queue is managed in this way. Before a new species is introduced to augment the ecosystem, one of the existing species is chosen at random, and saved into the current queue position. The queue is advanced, and the saved species in the next position is brought back to life. A random sample of 10 members of each species is saved in each queue position, allowing some diversity to be retained from the original population.

At the start of the simulation there are no saved creatures. Before the circular queue has completed its first cycle, a random creature is therefore used as a species seed. Once the queue is full, only 10% of species seeds are randomly configured. The remaining 90% of all introductions are species recalled from the saved data.

When recalling a cached species the 10 individuals are re-introduced as a new species in close proximity to each other at a random position in the space. They continue in reproductive isolation from all existing species to prevent the damaging crossover described earlier.

In this approach, even the total dominance of a new strategy over existing species will not lead to a single converged population. The queue may become filled by a particular form of creature, which has wiped out all other contenders. However, the sexual isolation of re-introduced species allows them to diverge anew, introducing distinct modifications to the dominant creature strategy, which may complement each other, filling different niches in the ecosystem.

GENOTYPE TO PHENOTYPE MAPPING

The genotype to phenotype mapping employed in the Black Shoals project is intended to address specific problems in the description of coupled systems of components.

Many approaches to the description of a phenotype depend upon restrictions in the final form adopted by the individual. In the case of robotics, a fixed bodily structure, and pre-defined sensory modalities are often assumed. In evolved neural networks, the topology of the network is often pre-defined. Restricting the possible configurations in this way did not satisfy the Black Shoals artists desire for open-endedness. Satisfying this criteria presented some problems, and these are not only due to the massive phenotype space involved.

Finding a general form of genetic expression for unknown couplings between multiple components of unspecified types is difficult. Many forms of evolutionary algorithm employ ‘tree structured’ genomes; directed acyclic graphs, which can be expressed in a linear form like a conventional genotype. However, this would exclude many forms of bodily structure with mechanically interesting properties – e.g. triangles, which are cyclic. It would further prevent the creation of recurrent, (cyclic), neural nets which are capable of memory - retaining state between inputs (in contrast with conventional feed forward nets). Finally, it is not obvious how tree structured genomes should be recombined (on mating) in a way which retains consistency in the phenotype.

In developing the encoding, the objective was to allow open-ended combinations of elements, with minimal assumptions from the designer as to the ways in which they may be coupled. This was intended to mirror the ways in which biochemical elements are combined to create real world organisms.

A generic method of encoding directed graphs from elements with predetermined dependencies was developed, as described below. This permitted further elements to be introduced without modifying any of the genetic operators to accommodate new elements.

Finally, the encoding had to permit the evolutionary process to control the number and type of phenotypic elements, and hence the complexity of their phenotype. Although the genotype was of fixed length, providing a maximum limit on complexity, genes were able to control each others’ expression or suppression through epistatic interactions, and hence determining the complexity of their phenotype.

These features of the encoding were very important, since at the beginning of the development effort, it was not known exactly what components would be developed, what their interrelationships would be, or which combinations of these elements would create viable creatures, making flexibility very important.

DEPENDENCY

In order to make a contribution to the fitness of the organisms, individual components are dependent upon each other’s existence to create stable modules.

In turn, these modules are dependent upon other modules to create a coherent whole. In our system, for example, a light elastic rod needs to be associated with two masses, one at each end, in order to be properly formed; A neural link needs a neuron at each end. Sensors and Actuators, although they pass information in different directions, both require an association with neurons and with physical body parts, coupling the physical body and the brain.

The final list of elements included in the final simulation is shown below, along with their descriptions and dependencies.

These represent the lowest level of interdependency and are enforced by the developmental process. Elements which fail to satisfy the dependencies detailed in the right hand column are incomplete, and are not included in the final phenotype of the creature. Elements which depend on incomplete elements are themselves incomplete.

The developmental process of the individual begins with a genome, containing several loci. Each individual locus describes a specific creature element. The information contained in the chromosomes maps elements into a developmental space within which dependencies are resolved.

The Developmental Space

The use of a developmental space provides a form of soft addressing, helping to negotiate dependencies. Individual elements are not specifically targeted by their location in the genome, but rather by the position they adopt in the space, which is itself genetically encoded, allowing couplings to form and dissolve through evolution. The development process need not be significantly disrupted by the re-ordering, removal or addition of individual loci.

At present, the number of loci in the genome is fixed for each species, but there is no reason why this should not vary given appropriately modified mutation and recombination operators, allowing increasing or decreasing complexity within a species. Since the position of a given element in the genome does not define its interpretation as part of the final configuration, such changes need not disrupt the interpretation of other elements.

Perhaps most importantly of all, the developmental process provides a means for the construction of arbitrary configurations – directed cyclic graphs of the form shown in figure 1 above.

Locuses and Elements

A locus contains the following information ;
i) type of element specified
ii) coordinate(s) defining its position
iii) coordinate(s) defining the area of search for dependent elements
iv) numerical parameters specific to the element

The generality of this format allows the process of interpretation of a locus to be agnostic to the specific elements included in the phenotype.

Elements are self-describing, in the sense that the number of co-ordinates required, the numerical parameters and their valid ranges may be ascertained simply by knowing the element type. It is therefore possible to create a random locus for an element type, and to implement operators for mutation and recombination which are applicable to arbitrary elements.

This permits the indefinite extension of the repertoire of elements without modifying the developmental process or the evolutionary algorithm.

CHROMOSOMES

In the present implementation, the genome is separated into two chromosomes, each containing 30 locus positions.

The first chromosome describes elements which may adopt a position in the developmental space independently of the others. These Positionables are the Masses (components of the physical body) and the Neurons (components of the brain). Their insertion into the space provides the basis for other elements to form initial associations and acquire a position within the space themselves. Although their participation in development is passive, acting only as targets for other elements, they may require a specific relationship to be formed with them in order to be complete. For a Neuron or Mass to participate in the phenotype of a creature, it must be connected to at least one other part of the creature. They must therefore be targeted by at least one higher level element, which is itself complete.

The second chromosome describes elements known as Element Sets. These Element Sets are the Rods, Neural Weights, Sensors and Actuators. These are elements which pursue associations amongst the existing elements in the developmental space according to individual search algorithms and satisfaction criteria. In practice, the same approach is adopted by all Element Sets. For each association, the locus contains a search co-ordinate, and a maximum radius of search, defining a circle from which appropriate elements may be drawn. An Element Set defines satisfaction criteria for each association. The nearest element of the appropriate type found in the space becomes the target of that association.

The position of an Element Set within the developmental space is defined as the mean of its search co-ordinates. The same neighbourhood search can therefore be used to target Element Sets in the same way as Positionables.

Once the process of development has created a graph of complete units, the creature is introduced into the domed space, with a random orientation, and from then on must fend for itself according to the energetic constraints described in the last section.

SUMMARY : GENOTYPE TO PHENOTYPE MAPPING

The genotype to phenotype mapping is capable of creating complex complete graphs of coupled components. The mapping has some key properties.

ß Epistasis : The expression of specific elements of the genotype depends upon epistatic interactions with other genes. This enables the retention of complex modules as dormant genetic material in the population, which can become re-expressed later and potentially combined with other modules. This characteristic is found in the homeobox genes of real organisms, enabling the reuse of genetic complexes for eye and limb formation, common genes which are shared between flies, human beings and mice. The development of such complexes is considered to have been key to the explosion of diversity in the Cambrian era. The negotiation of dependency between expressed elements provides a basis for epistatic interactions between genes.

ß Redundancy : Modules which are key can be re-expressed many times in the genotype of a specific organism, since the position in the genotype is not important. This ensures that these key components are not lost owing to mutation.

ß Neutrality : Many changes in a genotype make no change in the phenotype, either because they remain unexpressed, (through epistasis) or because they do not change the couplings between components, (for example when a coordinate changes by a small amount, and the associations formed in the developmental space remain identical). This encourages diversity of dormant genetic material within the population, and supports the development of new modules through the crossover of distinct material between two members of the same species.

RELATED WORK

This was not a benchmark problem, so it is difficult to compare results with other work in the field of evolved embodied organisms, although there are notable similarities.

Examples of embodied creatures created by artificial selection include Karl Sims seminal work using Thinking Corporation’s high powered Connection Machine1 2, Jeffrey Ventrella’s Gene Pool and Darwin’s Pond 3 and the Framsticks project 4. Other approaches from which inspiration was drawn were those of Theo Jansen and his Animaris Sabulosa 5, and Grey Walter’s Machina Speculatrix 6 7 8 This led to the christening of the creatures’ Animaris Speculatrix - speculating animal - which seemed appropriate to their interactions with the financial markets.

This piece of work combines many many of the characteristics of these projects, but distinctive aspects of the evolutionary approach can be identified, principally in the species management, the developmental process, and the element-based architecture. Limited space prevents me from representing these projects in the detail they deserve and forbids a detailed comparison.

RESULTS

This is a somewhat unconventional paper, with unconventional results, since the judgement of art is principally qualitative rather than quantitative.

The primary results are discernable with the naked eye. Creatures evolved which could coordinate muscular stretching and relaxing with foot control, leading to energy-efficient, natural looking gaits which allowed them to explore and discover pockets of trading activity within the star space. Some creatures developed sensorimotor control, responding to food in their environment by modifying their gait, turning, and heading towards food. A large variety of strategies were seen to emerge.

We were unable to determine if any creatures were able to predict the zones of trading activity. That secret has gone with them to their graves; The population cache was deliberately finite in order to avoid system failure when the disk was full, so no data exists on the many thousands of species who lived and died in the dome.

The results largely remain in the minds of the visitors to the exhibit. The Black Shoals project ran for three months at the Tate Gallery and, it is hoped, will tour the major cities of Europe.

FURTHER WORK

It is hoped that the Java™ implementation of the ecosystem in Black Shoals can be made available via the web as an applet, allowing web clients to watch the dramas of creatures on the desktop. Collaborators and investors are welcome to contact me if they are interested in supporting such a project.

REFLECTIONS

In designing the world of evolving creatures for the Black Shoals project, the objectives were clear. It was vital that the viewers were able to empathise with the creatures and their small existence.

The creatures’ relationship with their artificial world of stars is a mirror image of our relationship with the financial markets – they strive to survive, competing with each other in a world whose complexity they are too simple to fathom. To encourage the viewer to identify with the creatures, lifelike and purposeful behaviours had to emerge.

I believe that we achieved our objective. I was no less fascinated in the creatures which were brought forth by this artificial world, despite my intimate knowledge of their innards. Even while test-running the system in early trials, we all developed our favourites, egged them on to find food and survive, and were disappointed when they were eliminated by leaner, meaner creatures.

By contrast with the processes of computer animation, we did not control the creatures directly, but rather created a context in which their development could take place.

Achieving a lifelike quality therefore depended upon creating conditions which reflect those of nature; A system based upon Newtonian physics, constrained by limited resources, in which you have to compete to survive.

This paper is intended as a summary of how it was done, for the other geeks who like to play god.

Acknowledgments

Thanks to Josh Portway and Lise Autogena for inviting me to participate in the project and for all their input. Thanks to all the other contributors. Thanks to the curators and staff at the Tate Gallery for not throwing us out after dark, to let us get everything working for the opening. Thanks also to my management for supporting my involvement in parallel with my full time work for BTexact’s Future Technologies Group

Cefn Hoile is a researcher at BT Future Technologies Group.

[1] "Evolving Virtual Creatures" K.Sims, Computer Graphics (Siggraph '94 Proceedings), July 1994, pp.15-22.

[2] "Evolving 3D Morphology and Behavior by Competition" K.Sims, Artificial Life IV Proceedings, ed.by Brooks & Maes, MIT Press, 1994, pp.28-39.

[3] Jeffrey Ventrella‘s Website http://www.ventrella.com

[4] "Framsticks: towards a simulation of a nature-like world, creatures and evolution." M. Komosinski, Sz. Ulatowski, Proceedings of 5th European Conference on Artificial Life (ECAL99), September 13-17, 1999, Lausanne, Switzerland, Springer-Verlag (LNAI 1674), 261-265.

[5] Videos and commentary of a documentary about Theo Jansen online at http://carol.wins.uva.nl/~krose/animaris/

[6] Walter, W. Grey, "An Imitation of Life," Scientific American, May 1950, p42-45.

[7] Walter, W. Grey, "A Machine that Learns," Scientific American, August 1951, p60-63.

[8] Grey Walter Documents and Biography online at http://www.epub.org.br/cm/n09/historia/documentos_i.htm