Two more techniques are used to improve the performance of the algorithm:
- Fitness shaping – objective ranking: We discussed this technique previously. It's very simple. Instead of using the raw returns to compute the update, a rank transformation is used. The rank is invariant to the transformation of the objective function and thus performs better with spread returns. Additionally, it removes the noise of the outliers.
- Mirror noise: This trick reduces the variance and involves the evaluation of the network with both noise and ; that is, for each individual, we'll have two mutations: and .