Researchers at the Uber AI lab have proposed a coding scheme that uses the seed of a pseudo-random number generator to encode the phenotype ANN. In this scheme, the genome is represented as a list of seed values, which is applied sequentially to generate the values for all the weights (trainable parameters) of connections that are expressed between the nodes of a controller ANN.
In other words, the first seed value in the list represents the policy initialization seed, which is shared between the genealogy of the descendants of a single parent. All the subsequent seed values represent the specific mutations that are acquired during the evolution process by the offspring. Each seed is applied sequentially ...
