Summary
Neuroevolution applies evolutionary search to neural networks. Instead of training a network only through gradient descent and labelled examples, a population of candidate networks is evaluated, selected, and varied over generations. Depending on the design, evolution can target inputs, topology, learning parameters, or just the connection weights. (Brabazon, O’Neill and McGarraghy, Natural Computing Algorithms, see source-natural-computing-algorithms; Yannakakis and Togelius, Artificial Intelligence and Games, see source-ai-and-games)
Key ideas
- Works well when the success signal comes from full agent performance rather than labelled training pairs.
- Can evolve:
- input selection
- hidden-layer structure
- connection patterns
- connection weights
- Useful when the objective is non-differentiable, delayed, or simulation-based.
In practice
Game uses:
- evolving controller policies
- training agents for movement or tactics through whole-match performance
- searching network structures for specific game tasks
Evidence
- Natural Computing Algorithms highlights neuroevolution as a response to the huge combinatorial search space created by network architecture and parameter choices.
- Artificial Intelligence and Games treats neuroevolution as one of the notable hybrid methods joining evolutionary search with machine learning.
Implications
- Neuroevolution sits naturally between genetic-algorithms and machine-learning-games.
- It is especially relevant when a designer can simulate an agent repeatedly and assign a fitness score, but cannot easily define a supervised dataset.
Open questions
- Which small game benchmark would make the clearest classroom neuroevolution example in Unity?
- When should a student prefer standard reinforcement learning over neuroevolution?
Related
genetic-algorithms · machine-learning-games · particle-swarm-optimisation · source-natural-computing-algorithms · source-ai-and-games