摘要
本文提出了一个启发式规则划分办法并分析了并行规则推理的代价,设规则集合被划分为π=(π_1,π_2,…,π_k),则对于任何一个规则R_i∈π_i,在π_(i-1)中至少有某个R_i,R_i与R_i在某种代换下相关,如果串行推理对规则进行一次遍历所费时间代价为C,对于同样规则划分π进行多处理机或进程并行推理一遍遍历代价至多为C′,C′≤(k/~#R)C。其中~#R是规则集合R的基数,同时给出一个实例和推理代价的实验数据。
A heuristic rule set partition was presented in this paper. The partition is based on rule unification under some substitutions. Define a partition as follows: π= (π1, π2, …πk) For each rule in πi is unified with at least one rule in the πi-1, where i=2 to k. According to the partition, we implemented a concurrent forward chaining and backword chaining system with Object-Oriented Programming Language (Smalltalk- 80) on Sun station. For rule execution sequence E=(e1, e2, …, et) the cost of executing ei would be: C(ei) But the concurrent rule driving cost for the sequence will be C' = (K/#R)C(ei) where #R is the number of rules. We compare concurrent rule driving with sequential inference, (16 rules and 7 facts) the result is 480 milliseconds against 4000--4600 milliseconds.
关键词
人工智能
专家系统
划分
并行处理
rule
parallel processing
substitution
artificial intelligence/heuristic partition
processes
