Game-tree search plays an important role in the field of Artificial Intelligence （AI）. In this paper, we characterize one parallel game-tree search workload in chess： the latest version of Crafty, a state of art pr...Game-tree search plays an important role in the field of Artificial Intelligence （AI）. In this paper, we characterize one parallel game-tree search workload in chess： the latest version of Crafty, a state of art program, on two Intel Xeon shared-memory multiprocessor systems. Our analysis shows that Crafty is latency-sensitive and the hash-table and dynamic tree splitting used in Crafty cause large scalability penalties. They consume 35%-50% of the running time on the 4-way system. Furthermore, Crafty is not bandwidth-limited.展开更多
针对蒙特卡洛树搜索算法(Monte Carlo tree search,MCTS)收敛速度过慢,且在博弈过程中关键节点会出现信息丢失等问题,以中国象棋为载体,构建适用于中国象棋博弈系统的策略价值网络,提出了一种基于统计数据的并行蒙特卡洛树搜索算法(para...针对蒙特卡洛树搜索算法(Monte Carlo tree search,MCTS)收敛速度过慢,且在博弈过程中关键节点会出现信息丢失等问题,以中国象棋为载体,构建适用于中国象棋博弈系统的策略价值网络,提出了一种基于统计数据的并行蒙特卡洛树搜索算法(parallel Monte Carlo tree search based on statistics,SPMCTS)。将并行化的重点设置在MCTS四个步骤中最耗时的扩展和模拟步骤,有效避免了算法执行过程中的等待时差。并且引入一组新统计数据,这些数据用于在MCTS的选择步骤中修改节点的选择策略,保证在进行节点选择时获取和利用更多的可用信息,缓解信息丢失对精度造成的影响。实验结果表明,与现有并行蒙特卡洛树算法相比,SPMCTS在搜索速度上加快了约34%,且在对弈实验中,博弈胜率也能保持在80%左右。验证了SPMCTS的有效性。展开更多
文摘Game-tree search plays an important role in the field of Artificial Intelligence （AI）. In this paper, we characterize one parallel game-tree search workload in chess： the latest version of Crafty, a state of art program, on two Intel Xeon shared-memory multiprocessor systems. Our analysis shows that Crafty is latency-sensitive and the hash-table and dynamic tree splitting used in Crafty cause large scalability penalties. They consume 35%-50% of the running time on the 4-way system. Furthermore, Crafty is not bandwidth-limited.
文摘针对蒙特卡洛树搜索算法(Monte Carlo tree search,MCTS)收敛速度过慢,且在博弈过程中关键节点会出现信息丢失等问题,以中国象棋为载体,构建适用于中国象棋博弈系统的策略价值网络,提出了一种基于统计数据的并行蒙特卡洛树搜索算法(parallel Monte Carlo tree search based on statistics,SPMCTS)。将并行化的重点设置在MCTS四个步骤中最耗时的扩展和模拟步骤,有效避免了算法执行过程中的等待时差。并且引入一组新统计数据,这些数据用于在MCTS的选择步骤中修改节点的选择策略,保证在进行节点选择时获取和利用更多的可用信息,缓解信息丢失对精度造成的影响。实验结果表明,与现有并行蒙特卡洛树算法相比,SPMCTS在搜索速度上加快了约34%,且在对弈实验中,博弈胜率也能保持在80%左右。验证了SPMCTS的有效性。
