Evolutionary Algorithms are of the most successful algorithms in solving various optimization problems. Genetic network programming is one of the Evolutionary Algorithms with good capabilities in agent control problems. In this algorithm, the individuals’ structure is a directed graph. Using this structure, it is possible to model the solution of many complex problems. However, in this algorithm, crossover and mutation operators repeatedly break the structures of individuals and make new ones. Although this can lead to better structures, it may break suitable structures in elite individuals. Meanwhile, in stochastic environments, each time an individual is evaluated, it leads to different fitness values. So, calculating the fitness value of individuals requires evaluating each individual several times. This extremely decreases the evolution process speed. In this paper, inspired by mechanisms of ant colony algorithm, a new method is proposed to prevent the algorithm from iteratively breaking down the structures of individuals. This method improves the performance of individuals from one generation to the next using a constructive process. Unlike generative process that the individuals are generated by combination of some others, in constructive process they are produced according to the experience of previous generations. Using this mechanism, we not only prevent breaking suitable structures but also can manage uncertainty in stochastic environments. Our proposed method is used to solve two agent control problems when the environment is deterministic or stochastic. The results show that the proposed algorithm has very high ability in creating an efficient decision making strategies especially in stochastic environments.

进化算法是解决各种优化问题最成功的算法。遗传网络编程是进化算法中的一种，在代理控制问题上具有良好的能力。在这种算法中，个体的结构是有向图。使用这种结构，可以对许多复杂问题的解决方案进行建模。但是，在该算法中，交叉算子和变异算子反复破坏个体的结构并产生新的个体。尽管这可以导致更好的结构，但它可能会破坏精英个人的合适结构。同时，在随机环境中，每次评估一个人时，都会得出不同的适应度值。因此，计算个体的适应度值需要对每个个体进行几次评估。这极大地降低了进化过程的速度。本文在蚁群算法机制的启发下，提出了一种防止算法迭代破坏个体结构的新方法。此方法使用建设性的过程来提高个人从一代到下一代的性能。与生成过程不同，个体是通过其他个体的组合而生成的，而在生成过程中，它们是根据前几代人的经验生成的。使用这种机制，我们不仅可以防止破坏合适的结构，而且可以管理随机环境中的不确定性。当环境是确定性的或随机的时，我们提出的方法用于解决两个代理控制问题。