The Hitchhiker's Guide to Neuroevolution in Erlang
summary
Neuroevolution here implements DXNN in Erlang and claims concrete wins: topology and weight search are separated, topology evolves via random intensity mutation instead of crossover, and local search uses stochastic hill‑climbing to back up improved weight sets. Benchmarks show pole‑balancing and other tasks reaching solutions 3–4× faster than NEAT; a 250‑city TSP run converges to “good enough” routes. The system allows recurrent connections and uses typical GA knobs (eg crossover ≈0.7, mutation ≈0.005) and production Erlang processes: DXNN implemented in Erlang.
AI-written summary · about
chapters (26)
- 0:00Start
- 0:33Neural Nets
- 1:25Neurons
- 3:01Activation Function
- 3:31The Perceptron Learning Rule
- 3:53Calculate the Error
- 6:52A Binary Classifier
- 8:33Back Propagation
- 8:47Back Propagation Algorithm
- 10:43Neo-Darwinism
- 12:56Traveling Salesman Problem
- 13:43Genetic Operators
- 14:34Fitness Function
- 16:42Neuro Evolution
- 18:38Forex Trading
- 19:29The Genotype of a Neural Network
- 20:01Define the Genetic Operators
- 22:05Random Intensity Mutation
- 22:47The Mimetic Algorithm
- 24:14Mutant Allotment
- 24:35Offspring Creation
- 25:36Erlang
- 30:32Recurrent Connections
- 32:11Public State Steady State Evolution
- 32:48Benchmark for Neural Nets
- 34:28Benchmarks Results
related talks
30mEvolutionary Algorithms Perfecting the Art of "Good Enough"
46mTight Genes: Intro to Genetic Algorithms
48mTight Genes: Intro to Genetic Algorithms
57mSolutions That Evolve: Building Self-Improving Systems with Genetic Algorithms
37mSave the Princess
46mTight Genes: Intro to Genetic Algorithms
57mDarwinian Machine Learning: Principles of Machine Learning in Evo-devo, Evo-eco and Evolutionary...
15mEINNET: Optimizing Tensor Programs with Derivation-Based Transformations