- Start with a fully connected, but minimal, topology.
- Start with x-axis seed nodes.
- Choose a higher mutation rate for individuals which have a worse fitness than the avergage.
- Adapt the mutation rate according to the average behavioral distance
- Weight mutation: Change until the behavioral distance to the original individual changes by some percentage.
- Implement a Conrod GUI for experimenting with setting configuration options during the simulation run.
- Embed a RNG into every Genome.
- Record statistics, like number of mutations.
- Experiment with several different graphs.
- Make probability of structural mutation dependent on the complexity (number of nodes, number of links) of the genome.
- Substrate: Different placement
- Make weight mutation probability dependent on the current generation
- Make structural mutation dependent on the average node degree. For example, if there is a low connectivity of nodes, adding a new node is not a good thing.
- Add symmetric links, which, when updated, also update their counterpart.
- Add a fourth objective: Mutation work, which describes how much mutation has happened since the beginning for that individual.
- When adding a link, use a fixed weight for the second link
- The CPPNs we use, sum all inputs. This way, we cannot
represent e.g.
sin(x) * sin(y). Add aggregation functions/nodes, which can specify arbitrary functions on the inputs.