On the Potential for Open-Endedness in Neural Networks

Area of Science:

• Artificial Intelligence
• Evolutionary Computation
• Complexity Science

Background:

• Natural evolution drives increasing diversity and complexity.
• Artificial open-endedness requires methods that mimic evolutionary dynamics.
• Machine learning (ML) and artificial intelligence (AI) are often seen as too narrow for exploratory dynamics.

Purpose of the Study:

• To bridge the gap between evolution-inspired and ML approaches for creating open-ended systems.
• To review and address common barriers to open-endedness in evolution-inspired methods.
• To demonstrate the transferability of solutions between evolutionary and ML paradigms.

Main Methods:

• Reviewing barriers to open-endedness in evolution-inspired approaches (e.g., diversity collapse, complexity saturation).
• Mapping these barriers to analogous challenges within ML.
• Discussing the porting of solutions from evolutionary computation to ML.

Main Results:

• Identified common barriers to open-endedness in both evolutionary and ML systems.
• Demonstrated that solutions for evolutionary barriers can be adapted for ML.
• ML formulations offer new perspectives for resolving open-endedness challenges.

Conclusions:

• Researchers can use evolutionary and gradient-descent-based ML methods interchangeably.
• This interchangeability facilitates the design and creation of open-ended artificial systems.
• Synergistic use of these methods advances the pursuit of artificial general intelligence.