Virtual animals
Virtual animals (VT) are simulated multi-agent systems in an artificial biological environment. They are used to investigate emergent properties on the basis of neurobiological principles.
The starting point is an experiment for which an underlying internal model (working theory) exists that adequately describes the behavior of the object of investigation. This can be, for example, a hypothesis on the locomotion mechanism of a jellyfish or a neural model for integrating light / tactile / vibration events. In particular, models that use parameters that are not directly accessible in the experiment are suitable.
A corresponding virtual environment is created for the VT , which implements the essential model parameters. The environmental parameters can now be freely selected in the virtual experiment. The proactive virtual animal interacts with itself and the environment, depending on the experiment also with other virtual animals.
In comparison to the corresponding real experiments, the quality of the underlying model can now be checked using "normal operation". Furthermore, behavior in borderline areas can also be researched that would not be possible in real experiments (in practice or only under ethically intolerable circumstances). A quality criterion of the model can apply if the virtual animal has emergent properties that the real counterpart also shows.
While artificial life also uses an algorithm to calculate biological and social processes (for example, the solution of the traveling salesman problem with ants), virtual animals are based on connectionist neural networks . In addition to classic associative network architectures, learning through reward and punishment, hierarchical neural concept formation, body intelligence , adaptation , facilitation , depression and fractal structure formation may be used.
Virtual animals are bottom-up oriented. It is not only the performance in terms of specific objectives (such as foraging for food or flight behavior) that is analyzed, but also possible incorrect associations. Minimal changes in an existing network structure, for example simulating the partial death of neurons, can lead to unexpected statistical properties. The design of the network structure is not specifically specified. The neuron model used is ideally close to biology and operates on local rules such as Hebb learning , self-organizing maps or structural plasticity .
With regard to evolutionary development, nervous systems initially only perform very simple and reflex-like functions that are functional for survival. With increasing cerebralization and cephalization of the nervous system, the internal neuronal representation becomes more complex and allows more flexible adaptation to different environments and ecosystems ( selection pressure ). Examples of virtual animals are pheromone-oriented ants , termites , aplysia , worms , molluscs and other invertebrates as well as simple vertebrates . The concrete implementation is often more abstract and simulates more general problems such as foraging for food, simple classifications, exploration and social behavior .