Local evolvability of statistically neutral GasNet robot controllers

Area of Science:

• Evolutionary Computation
• Artificial Intelligence
• Robotics

Background:

• Evolutionary algorithms (EAs) are powerful optimization techniques.
• Understanding the dynamics of evolutionary search, particularly during neutral fitness periods, is crucial for improving EA performance.
• The concept of local evolvability offers a new lens to analyze the structure of evolutionary search spaces.

Purpose of the Study:

• To introduce and apply the concept of local evolvability to analyze population behavior during evolutionary search.
• To investigate the evolution of GasNet neural network controllers for a robotic visual discrimination task.
• To examine how local evolvability and solution robustness change over evolutionary runs.

Main Methods:

• Application of local evolvability metrics to analyze the search space.
• Focus on the evolution of GasNet neural network controllers.
• Analysis of evolutionary runs exhibiting neutral fitness epochs and population takeovers.

Main Results:

• The evolutionary process frequently encounters long periods of neutral fitness.
• Local evolvability properties of the search space fluctuate during population takeovers but do not increase further after takeover.
• No evidence of neutral evolution for increased solution robustness was found.

Conclusions:

• Local evolvability is a dynamic property of the search space that varies with population dynamics.
• The lack of increased robustness evolution may be attributed to EAs focusing on areas with inherent robustness.
• Further research into the interplay between local evolvability, robustness, and EA strategy is warranted.