基于蒙特卡洛树搜索的土石方动态调配算法及验证

Earthwork dynamic allocation based on the Monte Carlo tree search algorithm and its verification

针对堆石坝填筑进度控制以及土石方动态调运问题,受AlphaGo-Zero的启发,本文提出了一个基于蒙特卡洛树搜索(Monte Carlo tree search,MCTS)的土石方智能动态调配模型。该模型以当前累计填筑工程量、紧邻前一月份完成工程量以及当前月份为状态,用各月填筑工作面对应的填筑可达强度约束动作空间,综合考虑节点工期、总工期、坝面施工机械费用和土石方调运费用等因素构造奖励函数。此外,结合本文研究问题的特点,对MCTS迭代中的上限置信区间算法(upper confidence bound apply to tree,UCT)进行了改进和比较分析,最后以一个工程实例对本文提出模型的有效性进行了验证分析。结果表明,与施工仿真相比,以MCTS为框架的土石方动态调配模型的计算分析时间大大减少,为土石方动态调配问题提供了新的模型与手段。

This study aims at the problems of filling schedule control and earthwork dynamic allocation of a rockfill dam.Inspired by AlphaGo-Zero,an intelligent earthwork dynamic allocation model(EDAM)based on Monte Carlo tree search(MCTS)is presented.The state of EDAM is divided into the current cumulative filling volume,the filling volume last month and the current month.The action space of EDAM is restrained by the upper limit of filling intensity,which is decided by the filling working face.The reward function of EDAM is constructed by the joint construction period,the total construction period,the cost of construction machinery on the filling working face and the cost of earthwork allocation.In addition,the upper confidence bound apply to tree algorithm(UCT)in MCTS iteration process is improved and examined based on the specific problems studied.An engineering example is ultimately used to verify the feasibility of the EDAM.The results show that the calculation and analysis time of the EDAM is dramatically reduced,compared with the construction simulation,and a new model and means is provided for the earthwork dynamic allocation.