Q&A 36.3 — NeuroevolutionEngine: NEAT Configuration & Topology Questions

[web3guru888]

NeuroevolutionEngine — Q&A Thread

Use this discussion for questions about Phase 36.3 NeuroevolutionEngine.

Common Questions

Q: How to set NEAT compatibility threshold?

• Start with δ_t = 3.0 (Stanley & Miikkulainen 2002)
• Use dynamic threshold: increase if too many species, decrease if too few
• Target 5-15 species for most problems

Q: When to use NEAT vs HyperNEAT?

• NEAT: small networks (<100 nodes), direct encoding sufficient
• HyperNEAT: large regular networks, spatial/geometric structure matters
• HyperNEAT excels at vision, spatial reasoning, game boards

Q: How to prevent bloat in topology search?

• Real et al. (2019): aging regularization — older genomes penalized
• Parsimony pressure: add complexity penalty to fitness
• NEAT's complexification naturally starts minimal

Q: ES vs NEAT for weight optimization?

• ES (Salimans et al. 2017): fixed topology, massive parallelism
• NEAT: co-evolves topology + weights, sequential generation model
• Hybrid: use NEAT for topology, ES for weight refinement

Related issue: #741
Planning: #738