自适应遗传算法学习1--crossover设计

crossover operater plays important role in GA algorithm, it is used to generate new solutions. we hypothesize that the best crossover operator is likely not only depend on the particular packing problem, the population size and other parameters of GA, but also on the current state of the population pool. With this respect, we select amount of top best solutions as the population pool and apply two different crossover operators, which are preserve-assignment crossover operator(cr1) and probabilistic-uniform crossover operator(cr2), both of operators will be demonstrated in the following section,
each operator takes two parents and produces two children. Since both of operators have different aims, adaptive crossover scheme is designed to give more chance to the wellperformed crossover in the evolution process.

 

we propose a adaptive crossover scheme that generates offsprings using some probability distribution over the set {cr1, cr2}. That is, offsprings are either generated using cr1 or cr2 at each generation. Initially equal probabilities are assigned to both crossover rules and these probabilities are updated according to some rules based on reward (for being successfully) and penalty (for being unsuccessfully). A probabilistic adaptation decides which crossover operator to use most in generating children. The adaptation procedure allows the algorithm to bias the crossover that solves a given problem in an efficient and robust manner.

 

The probabilistic adaption follows the concept of a learning system used in learning automata.The reward and penalty strategies demonstrated by reinforcement-learning scheme have been proposed, both linear and non-linear schemes have been applied in different problems successfully.

 

Initially, equal probabilities (pcr₁₁ , pcr₂₁ ) = (0.5, 0.5) are assigned to each of thecrossover operator cr1 and cr2, and the probabilities are updated according to the reinforcement learning scheme, at each iteration depending on whether the operator used performs relatively successfully or not. The probabilities pcr1k and pcr2k , where pcr1k +pcr2k = 1, are updated according to the prior performance and the current performance of the crossover, if the crossover performs more successfully than its counterpart in current iteration, the probability is increased using the rule (12), where 0 2 (0, 1) is a reward coefficient.

pk = pk−1 + 0pk−1(1 − pk−1)

otherwise, the probability will be reduced by the rule (13), where 1 2 (0, 1) is a penalty
coefficient.

pk = pk−1 − 1pk−1(1 − pk−1)

we set pcr1k to cr1, because pcr1k + pcr2k = 1, we just need to update pcr1k using (12) and(13), sequently attain pcr2k by pcr2k = 1 − pcr1k .

 

There is another important part is to determinate when to reward/punish the crossover operator, the strategy is determined by the computational expriments.