afcarl / neat-suckgeun

Implementation of NeuroEvolution of Augmenting Topologies (NEAT)

Overall Work Flow

preparation

1. create 100 neural networks with 50 inputs and 10 outputs; no hidden layer
2. create speciation records (currently just one; initial species with no hidden layer)

test neural networks

1. feed forward, get result

analyze results

1. evaluate fitness of each neural networks with the results (use modified fitness function such that no one species overwhelm the entire population)
2. rank neural network in each group
3. drop 50% of lesser neural networks from each species.

get next generation

1. from each reduced species, first, mutate (check if mutation or crossover is first)
2. then crossover
3. global innovation number, connection mutation, node mutation, weight mutation
4. E, D, how to crossover
5. inter species crossover?

speciation

1. speciate the neural networks using the "compatibility distance" function
2. choose one genome from "previous" generation
3. compare that to current generation
4. if compatibility threshold satisfies, assign to the species
5. if no species exists, create a new species

repeat the process

1. go back to "test neural networks"

---

Details

Feed forward

1. define function
2. set it as main function to neural network using getattr
3. calculate the inputs then output it.
4. feed forward
5. get result

initializing Neural Networks, History, and innovation counter

1. create initial connect gene for one neural network using workplace info
2. when creating, increment history and counter
3. assign value to nns, history, and counter

Adding new connection

1. Add_connect(in, out)
   1. does connection exists in global history?
   2. Is valid connection?
      1. is in-in connection?
      2. is out-out connection?
      3. is recursive connection?
      4. does nn have it?
      5. is it in global history?
   3. If all the above are not true,
      1. increment global innovation counter
      2. add the connection info to individual history