Measure recovery performance of a feature-cut creature

[benelot]

Find out about generalization properties of neural networks. Evolve morphology with simple controllers until it walks. Sweep the controller, add a learning controller and learn again how to walk. If it can walk, cut off features and see if it can learn how to walk again faster than a new randomly initialized controller. Generalize this behavior over multiple environments.

• Initial performance drop because over missing feature
• Recovery of learning creature to maximize fitness function again.
• Variate the feature mutation and variate the terrain and the creature subject to generalize
• The variation among the variated parameters is the confidence of the generalization.