An evolutionary algorithm that constructs recurrent neural networks

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Standard methods for recurrent neural network (RNN) induction impose architectural constraints, limiting task adaptability.
• The complex interplay between RNN structure and function necessitates flexible induction approaches.
• Evolutionary computation offers powerful search capabilities for complex optimization problems.

Purpose of the Study:

• To address the limitations of current RNN induction methods.
• To propose and evaluate an evolutionary program for simultaneous structure and weight acquisition in RNNs.
• To enable the emergence of complex behaviors and topologies not typically captured by standard methods.

Main Methods:

• Developed GNARL, an evolutionary program employing evolutionary programming for RNN acquisition.
• Utilized a population-based search strategy to simultaneously evolve network structure and weights.
• Contrasted the proposed evolutionary approach with standard genetic algorithms for network induction.

Main Results:

• Demonstrated the inapplicability of genetic algorithms for network acquisition tasks.
• Showcased GNARL's capability to simultaneously acquire both structure and weights of recurrent networks.
• Highlighted the emergence of complex behaviors and topologies through GNARL's empirical acquisition method.

Conclusions:

• Evolutionary programming, as implemented in GNARL, is a suitable method for recurrent neural network induction.
• GNARL overcomes the architectural limitations of standard network induction techniques.
• The proposed method facilitates the discovery of novel and complex network architectures and behaviors.