Evolution of cartesian genetic programs for development of learning neural architecture.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Evolutionary Computation

Background:

• Artificial neural networks (ANNs) traditionally focus on synaptic weight adjustments, often overlooking the genetic basis of neural functions.
• Evolutionary approaches in ANNs have primarily concentrated on synaptic plasticity, neglecting the structural and functional evolution of neuronal components.

Purpose of the Study:

• To explore evolutionary computation for discovering biologically analogous computational functions for neuronal sub-components.
• To demonstrate that incorporating biological plausibility, such as evolving neuron morphology, can lead to intelligent behavior.
• To investigate the learning capabilities of a novel, evolutionarily developed neural network model.

Main Methods:

• Utilized Cartesian genetic programming (CGP) to evolve computational functions for neuronal components (soma, dendrites, axon branches).
• Developed a compartmental neuron model where components can grow or die, altering synaptic morphology during problem-solving.
• Evaluated the model's learning potential in single-agent (Wumpus World) and competitive multi-agent scenarios using CGP-controlled computational networks (CGPCN).

Main Results:

• The evolved neural network model demonstrated the ability to modify its structure and functions during problem-solving.
• Agents controlled by independent CGPCNs exhibited significant learning capabilities in both cooperative and competitive environments.
• The system successfully generated intelligent behavior through the evolution of neuronal sub-component functions and morphology.

Conclusions:

• Evolutionary computation, particularly CGP, can effectively derive biologically plausible neuronal functions and structures.
• Evolving neuronal morphology alongside computational functions enhances learning and intelligent behavior in artificial agents.
• This biologically inspired approach offers a promising direction for developing more sophisticated and adaptable artificial intelligence systems.