Artificial life simulators and their application

Abstract

The project is directed towards A-life simulation methods and their application. The aim is to create specialized simulators of different initial bases (Swarm system, Lindenmayer systems, evolutionary algorithms) which take into consideration the requirements of particular application fields. The core of the project will be the utilization of these simulation tools in solving a relatively wide scale of practical tasks:
monitoring the global ecosystem behavior, utilization of algae in eliminating heavy metals from water, monitoring the effect of signal spread in plants on their morphogenesis, data mining, prediction tasks and constraint problem solving.

Research group members:

Description of the project

a) Present state of the subject

Artificial life (Alife) is a general method that we can describe as generation of macroscopic behavior from simple microscopic cooperating elementsand this behavior can be interpreted as a manifestation of life (Bonabeau & Theraulaz, 1995). Artificial life follows two goals:

  • to enrich our knowledge about nature
  • to improve our artificial model concepts and by doing this to reach their better performance.
    • It is done by abstracting of basic dynamical principles that are followed by biological phenomena, by implementing of this dynamics in a different physical environment (e.g. in a computer) and by introducing the new types of experiments to this system. The novelty lies in the synthetic approach (i.e. effort to create global behavior on the basis of cooperating elementary primitives). This is different from a classical biological approach which is analytical (i.e. effort to decompose complex phenomena to partial components). These differences between artificial and natural systems are clear on the level of conformational and functional primitives. On the level of collaboration the contrast is less distinctive and finally the difference totally fades away on the level of synthesized processes, effects and structures.

    Proposed project plans to deal with this form of synthetic approach in software platform, which is usually divided to simulation and instantiation of a life processes. In the case of classical life process simulations the basis is in the set of differential or difference equations that express mutual relationships and behavior of the partial components (e.g. genes, alleles, individuals or species). The new simulation approach is based on the bottom-up synthesis. Firstly we create a population of data structures where each structure corresponds to a single entity. These structures contain variables defining the state of an individual. Furthermore we define rules defining how the individuals interact with one another and with the environment. As the simulation runs, populations of these data structures interact according to local rules, and the global behavior of the system emerges from those interactions (Ray, 1995).

    In the case of simulation the data structure represents a real living entity. In the case of instantiation the data structure does not represent anything else. We look at such a structure as liveone and again we observe how their local behavior and interactions will display on a global level. Outlined thoughts enable to state that artificial life does not have to be only sheer simulation of some manifestations of natural life but that natural life on the Earth and some artificial forms of life are both just different instances of LIFE itself.

    Artificial life simulators can be divided according to more aspects. One of them can be a complexity of population features that is achievable by the simulator (ranging from simple games of LIFE to complex simulators of the type like SWARM or TIERRA). The type of evolution process used in simulator can be another aspect. Two principles are applied here, the case when abilities, that are learned during a lifetime, are inherited from a parent to an offspring (Lamarckian evolution) through a genotype (here it is actually a phenotype) and the case when there is no inheritance of learned abilities (Darwinian evolution).

    Furthermore the simulators can be divided according to interactions between organisms in the model environment: there are systems with local or global interactions.

    The core of the complexity development in the life systems is the method of the ochastic information transferfrom the environment to the genome of adapting cell by a process of mutation. This is also implemented in many artificial life simulators.

    The TIERRA system is suitable for the experiments focused directly on the modifications of the organisms genetic code. It appears to be an advantageous feature for evolving the optimized algorithms or finding a new algorithmic problem solution. TIERRA does not require big operational memory or high performance processor. Unfortunately it does not enable interactive observation and evaluation of the simulations. They work for more than two years on the complex visualization environment with Tcl/Tk libraries but it is still not available and so TIERRA shifts to the area of tools that require more than average programming skills.

    The AVIDA system (Adami&Brown, 1994) represents an auto-adapting genetic system with local interactions between the population individuals and with casual propagation of the information on two dimensional areas. Parallelism of the system is ensured by actualization mechanism that enables the string execution in the random order (as it is in the array of cellular automata). Local interactions lead to more diverse population with bigger range of genotype and lower risk to settle in local extreme what happens if the populations too homogenous. From the evolvability dependence on population size and mutation rate it was found out that learning rate grows faster in the bigger worlds and so it offers wider area for evolution. Upper limit of error catastrophe is universal for all population sizes. The flexibility of AVIDA enables its use in many applications. Besides of string reeding or executing the user defined functions it can be used also for research in evolution processes, ecology or immunology.

    Simulation system SWARM (Minar et al., 1996) appears to be the most complex and with the best simulation environment for the user including parameters setting and visualization of the experiments results. Very useful is also the use of objective language (Objective C or Java) as a source code for the organisms and it is transferable between various platforms. SWARM is in a permanent development, there are always new libraries added that enable direct approach to already programmed algorithms. The amount of user applications is growing and SWARM is more and more used for various problems solving with a help of artificial life

    There are projects from the area of various biological, ecological or sociological simulations. They are needed especially in systems of huge complexity e.g. a planetary ecosystem (Lovelock, 1995), because real system cant be studied for the reasons of long time scale (some processes last for years or thousands of years) or a dimension reasons (huge amount of biological species). The computer model enables also manipulations with a model that would not be possible in reality e.g. catastrophes. #The majority of plant modeling efforts was focused on creating realistic pictures of plants so far (Hogeweg & Hesper, 1974). The testing of a hypothesis concerning the mechanisms of a plant morphogenesis and the examination of the relationship between these mechanisms and the evolution of a morphogenesis was put aside (partly because of low interest from the biologists). The only exception is probably the classical Prusinkiewiczs example with the plant Mycelis muralis flower topology where it is possible in the model to test the hypothesis whether the signals assumed from biological data are sufficient for the creation of a realistic model or more signals are needed (Prusinkiewicz & Lindenmayer, 1990).

    There are not many Alife applications for solving the real technical tasks although the approaches based on evolutionary computations are widely used for solving various optimization problems (Fayyad et al., 1996). There is a possibility of using the described Alife principles for the tasks with high computational complexity, e.g.: the tasks within the knowledge discovering process in large databases especially data mining algorithms znalostm (Cios et al., 1998). As far as we know there are no published results from this area.

    We can obtain an effective tool for real simulations and designing the Alife applications by using the means of distributed artificial intelligence as agent oriented systems. Large palette of agent-based technologies use enables to build an optimal application with an easy possibility of enlarging and modifications thanks to modular characteristic of agents systems in connection with component technologies.

    There is another promising area in the applications of Alife principles evolutionary algorithms. They enable to search a space of potential solutions in a very effective heuristic way. The population principle is used for this search which allows a parallel exploration. The EAs represent a modern tool usable for a wide range of technical or non-technical tasks.

    The visualization of these algorithms represents a hard problem because of huge amount of information and its multidimensionality. Despite this it is a very helpful tool for study, measurement and modeling of algorithms dynamical behavior. Visualization itself has different forms in dependence on examined characteristics. Approaches based on data set variation features are often used, especially the analysis of main components. Their goal is the extraction of significant global features of a data set and the use of these components for coding the data instances in a compact form. Another group of approaches is based on a distances between particular points and tries to implement a transformation that will preserve the original distances (Collins, 97) (Shine & Eick, 97).

    Evolutionary algorithms belong to a class of weak algorithms they are applicable for very wide palette of tasks (the task has to satisfy only some requirements e.g. its solution can be visualized as a point in a multidimensional space etc.). On the other side, they dont reach the effectiveness of algorithms specialized in particular type of problems. The solution is hybridization by domain specific knowledge or by the other algorithm types.

    b) Particular contribution expected

    For the main asset of the project we consider development of new simulations in SWARM system with focus on these main areas:

  •  theoretical model implementation with a concentration to study of deeper contexts in ecological and environmental theories (e.g. Gaia theory),
  •  analysis of the relationships individual environment, individual individual and analysis of the global behavior emergence (as environmental regulation, population stability and its adaptation to the environment),
  •  analysis of coordination technologies in multiagent environment of the type controlling agent agents or equivalent environment agent - agents,
  •  analysis and implementation of the living organisms models (bacteria, algae and viruses) with possibility of using the simulation instead of real biological experiments,
  •  analysis and implementation of the biochemical experiments over the modeled organisms and their comparison with real experiments,
  •  design and implementation of the cognitive agents models, development of libraries with basic agent types and tools for their generation and modification. The goal is to make agent generation process easier,
  •  implementation of standard component technologies CORBA and COM into the SWARM system and its utilization in distributed simulations in the various network topologies and analysis of their influence to time and memory demands of the simulations,
  •  design and implementation of own visualization technology over the distributed topological structure in multiagent environment.
    •  The use of Lindenmayer systems for modeling of the morphogenetical plant processes is the special case of a natural process simulation. In this area we expect following assets:
  • design of theoretical plant morphogenesis model enabling the observation of the parameters influence to the formation of specific morphologies,
  • model design of various tropism influence to plant morphogenesis,
    •  Another asset is expected in a field of Alife technique applications in the knowledge discovery: We expect the following assets in evolutionary algorithm area:
  • design of new evolutionary algorithm visualization methods for static and dynamic (variable number of dimension) search space,
  • interconnection between the visualization methods of the evolutionary algorithm activity and possibilities of interactive changes of the ew user defined visualization,
  • design of hybridized evolutionary algorithms for specific problem solving (e.g. prediction problems, constrain problems etc.)
    • c) Proposal of the ways how to reach the project goals

    We propose the following steps in the area of analysis, implementation and testing of the theoretical or biological model:

    In the area of plant morphogenesis research we assume these steps:
  • theoretical model development of a plant morphogenesis including the signal propagation ,
  • design and implementation of a flexible graphical program tool for plant morphogenesis, generation on the basis of Lindenmayer systems,
  • systematic study of a possible parameter space and checking whether certain parameter combination produce.
    • In the examination of Alife principles utilization for problem solving that is connected with knowledge discoveries in the databases we expect following procedure:
  • analysis of the agents features in context of knowledge discovery process in large databases,
  • knowledge discovering process analysis of the particular steps with respect to utilization of software agent features for their more effective support,
  • necessary attributes analysis of the primitives (status variables, interaction rules, fitness functions etc.) from the view of features emergence that are interesting for the knowledge discovery process.
    • We plan the following procedure in problem solving connected with evolutionary algorithms: References: