Neutrality, which is caused by highly redundant mappings from genotype to phenotype or from phenotype to fitness, has been recognized as an important factor for artificial genetic search. However, this research field has still been trying to overcome the difficulties of analyzing evolutionary dynamics at the level of phenotype space or fitness space for about ten years. On the other hand, considering that population genetics explains the change of gene frequency in a population, that is, dynamics at the level of genotype space, we expect that various techniques developed in population genetics might be useful for analyzing dynamics of artificial evolution. In this paper, we apply the Nei's standard genetic distance to artificial evolution. Several computer simulations are systematically conducted by applying a standard genetic algorithm to a tunably neutral NK landscape in order to clarify the characteristics of the Nei's standard genetic distance in genetic algorithms. The results prove several consistencies with the (nearly) neutral theory of molecular evolution.
A novel approach to Topology and Weight Evolving Artificial Neural Networks (TWEANN) is presented. This comes from the following two considerations. (1) Artificial evolution is designed without recombination which mostly generates an offspring whose fitness value is considerably reduced than those of parents. Instead of recombination, two types of topological mutations are provided as genetic operations. One is for inserting a new neuron and the other is for generating a new synaptic connection. The both of them are designed carefully to retain the current evaluation value. (2) A new encoding technique such that a string is defined as a set of substrings called operons is introduced. This is based on our idea that each potential function should be encoded into a different part of genetic information. In order to validate our approach, computer simulations are conducted by using one of the standard reinforcement learning problems called the double pole balancing problems without velocity information. The results obtained are compared with the results by NEAT, which is recognized as one of the most powerful techniques in this problem domain. It is found that our proposed approach yields competitive results, especially when the problem is tuned to be difficult.
Multi-agent reinforcement learning features many problems, one of these being state-space explosion based on combinations of policies that each agent has. Generally, if agents can share their policies, they can effectively search their enormous state space; that, however, simultaneously produces a risk of local optimality. Hence we propose a novel policy-sharing system based on the Learning Classifier System, on which agents locally share their policies. The aim of this system is to decrease the probability of falls into local optimality, and to effectively reduce state space by policy sharing. To verify the above, we use simplified soccer, which has discrete space and time.
We consider, in this paper, the capacitated multi-level facility location problem as a supply-chain network design. To cope with such large-scale and complicated problems, we have presented a practical method that divides first the original problem into two sub-problems, and then combine each result consistently. Each sub-problem, i.e., distribution side sub-problem and procurement one is solved likely by a hybrid method that employs meta-heuristic method like tabu search and graph algorithm for solving a transformed minimum cost flow problem in a hierarchical manner. Moreover, we have given a practical adjusting method for thus separately solved solutions associated with the Lagrangian relaxation. Through numerical experiments, we have confirmed that the proposed method can outperform a popular commercial software. On the average, the approximated rate of the obtained results was within 1.006 compared with the final solution from the commercial software.