Coevolutionary Learning Algorithm

From GM-RKB
Jump to navigation Jump to search

A Coevolutionary Learning Algorithm is an evolutionary algorithm in which the fitness of an individual is subjective.



References

2012

  • (Popovici et al., 2012) ⇒ Popovici, E., Bucci, A., Wiegand, R. P., & De Jong, E. D. (2012). Coevolutionary principles. In Handbook of Natural Computing (pp. 987-1033). Springer Berlin Heidelberg[1].
    • (...) The inspiration for coevolutionary algorithms (CoEAs) is the same as for traditional evolutionary algorithms (EAs): attempt to harness the Darwinian notions of heredity and survival of the fittest for simulation or problem-solving purposes. To put it simply, a representation is chosen to encode some aspects of potential solutions to a problem into individuals, those individuals are altered during search using geneticlike, variation operators such as mutation and crossover, and search is directed by selecting better individuals as determined by a fitness evaluation. With any luck the iteration of these steps will eventually lead to high-quality solutions to a problem, if problem-solving is the aim; or to interesting or realistic system behavior. Usually, EAs begin with a fitness function: a function of the form [math]\displaystyle{ f : G \rightarrow R }[/math] that assigns a real value to each possible genotype in [math]\displaystyle{ G }[/math]. Given such a function, the fitness relationship between any two genotypes [math]\displaystyle{ g_1, g_2 \in G }[/math] is clear: we compare [math]\displaystyle{ f(g_1) }[/math] with [math]\displaystyle{ f(g_2) }[/math] to see which is more fit. By contrast, CoEAs do not use such a direct metric of the fitness of individuals. Instead, two individuals are compared on the basis of their outcomes from interactions with other individuals.

2011

2007

  • (Design Decision Wiki, 2007) ⇒ https://wiki.ece.cmu.edu/ddl/index.php/Coevolutionary_algorithms Retrieved on 2017-06-04
    • (...) A coevolutionary algorithm is an evolutionary algorithm (or collection of evolutionary algorithms) in which the fitness of an individual is subjective; that is, the individuals are evaluated based on their interactions with other individuals. According to the nature of these interactions, coevolutionary algorithms fall into two main groups: Competetive Coevolutionary Algorithms and Cooperative Coevolutionary Algorithms. In the case of cooperative algorithms, individuals are rewarded when they work well with other individuals and punished when they perform poorly together. For example, consider an algorithm where each population represents a piece of a larger problem, and it is the task of those populations to evolve increasingly more fit pieces for the larger holistic problem. In the case of competitive algorithms, however, individuals are rewarded at the expense of those with which they interact. For example, consider a predator-prey model in which individuals in one population represent some kind of device (e.g., a sorting network) and individuals in another population represent some kind of input for the device (e.g., a data set), and the object of the first population is to evolve increasingly better devices to handle the input, while the object of the second population is to evolve increasingly more difficult inputs for the devices.

(2005)

2004

  • (Sim et al., 2004) ⇒ Sim, K. B., Lee, D. W., & Kim, J. Y. (2004). Game theory based coevolutionary algorithm: a new computational coevolutionary approach. Int J Control Autom Syst, 24, 463-474. [2]
    • Abstract- Game theory is a method of mathematical analysis developed to study the decision making process. In 1928, Von Neumann mathematically proved that every two-person, zerosum game with many pure finite strategies for each player is deterministic. In the early 50's, Nash presented another concept as the basis for a generalization of Von Neumann’s theorem. Another central achievement of game theory is the introduction of evolutionary game theory, by which agents can play optimal strategies in the absence of rationality. Through the process of Darwinian selection, a population of agents can evolve to an Evolutionary Stable Strategy (ESS) as introduced by Maynard Smith in 1982. Keeping pace with these game theoretical studies, the first computer simulation of coevolution was tried out by Hillis. Moreover, Kauffman proposed the NK model to analyze coevolutionary dynamics between different species. He showed how coevolutionary phenomenon reaches static states and that these states are either Nash equilibrium or ESS in game theory. Since studies concerning coevolutionary phenomenon were initiated, there have been numerous other researchers who have developed coevolutionary algorithms. In this paper we propose a new coevolutionary algorithm named Game theory based Coevolutionary Algorithm (GCEA) and we confirm that this algorithm can be a solution of evolutionary problems by searching the ESS. To evaluate this newly designed approach, we solve several test Multiobjective Optimization Problems (MOPs). From the results of these evaluations, we confirm that evolutionary game can be embodied by the coevolutionary algorithm and analyze the optimization performance of our algorithm by comparing the performance of our algorithm with that of other evolutionary optimization algorithms.

1998

  • (Windrum & Birchenhall, 1998) ⇒ Paul Windrum, and Chris Birchenhall. (1998). “Is product life cycle theory a special case? Dominant designs and the emergence of market niches through coevolutionary-learning.” In: Structural Change and Economic Dynamics 9, no. 1.