随机动态规划的混合算法研究

Study on the Hybrid Algorithm for a kind of Stochastic Dynamic Programming

针对随机条件下动态规划模型的主要特点,运用智能算法混合编程理论,设计了一种探索多阶段决策问题的智能混合算法。该算法首先将问题转化成一族同类型的一步决策子问题,然后利用随机模拟和遗传算法,依据训练样本形成的训练神经元网络,在单步决策中寻求最优策略和最优目标值,逐个求解,再据初始状态逆序求出最优策略序列和最优目标值。仿真结果表明,该算法具有一定的通用性,初始设计点可以随机产生,其计算精度不因函数的非线性强弱而受影响,对目标和约束的限制较少,可应用于多种形式的随机多阶段决策优化问题,较好地满足了随机动态规划模型求解和优化的要求。

An algorithm, which is used to solve a kind of the stochastic dynamic programming models, is proposed＇based on the theory of hybrid programming to explore the multi-stage military decision-making problem. The target problem is divided into a set of single-step problems, then the optimization policy and goal value of the sub-problem of single step is obtained by Random-Simulation and the genetic algorithm on the training sample of NN. The optimization policy and goal value of the problem are obtained step by step in inverted sequence from the initialization value. The simulation result shows that the algorithm meets the needs to optimize the stochastic dynamic programming models with the features as follows： be universal, the initial value be generated randomly, the precision be insensitive to the nonlinearity of the functions, be less restriction to the target function and the constrained conditions.