This is the main paper introducing the NEAT system. This system is a direct-encoding based way of dealing with neuroevolution (evolution of ANNs). The encoding is based on a genome sequentially specifying each of the connections between modules of the network. Several tickes are used to make it possible applying GA methods to evolve networks:
- Historical tracking of genes to be able to align architectures and mate them.
- Innovation protection with speciation. A notion of speciation is used, and fitness is shares among entire species to make innovations that aren’t immediately beneficial still possible.
The method is first tested on XOR and is further applied to a 2 arms pole stabilization problem. NEAT seems capable of achieving satisfying solutions quicker than most methods except maybe ESP ((Gomez, Miikkulainen 1999)).
This paper is very influential (one of the most cited one in neuroevolution). The idea is very interesting but doesn’t realy seem able to scale since the encoding of the architecture seems so direct. I belive HyperNEAT solves those issues.
Other than this, the main task solving is based on optimizing an objective with a GA.
- Kenneth O. Stanley, Risto Miikkulainen. . "Evolving Neural Networks Through Augmenting Topologies". Evolutionary Computation 10 (2):99–127. DOI.
- Faustino J. Gomez, Risto Miikkulainen. . "Solving Non-markovian Control Tasks with Neuroevolution". In IJCAI, 99:1356–61.