推理机制
-
forward chaining （正向推理）
-
backward chaining（逆向推理）
假设我们有下面四条规则：
1.如果X呱呱叫并且X吃苍蝇，那么X是只青蛙
2.如果X唧唧咋咋的叫并且X会唱歌，那么X是只金丝雀
3.如果X是青蛙，那么X是绿色的
4.如果X是金丝雀，那么X是黄色的
假设我们又得到两个事实：
- Fritz croaks （Fritz呱呱叫）
- Fritz eats flies（Fritz吃苍蝇）
一方面我们可以通过forward chaining 得出结论：
-
Fritz是只青蛙
-
Fritz是绿色的
另一方面如果我们的目标是证明Fritz是绿色的，那么通过 backward chaining 求解决会更有效率
1.在知识库里要证明Fritz是绿色的，Fritz就需要是一只青蛙
2.若要Fritz是一只青蛙，Fritz就必须呱呱叫而吃苍蝇，而这正好符合已知事实，所以Fritz是绿色的这个命题就被证明了。
看起来backward chaining 好像有点麻烦，不过当forward chaining有太多前进路径，采用backward chaining就会非常高效了，这有点像函数式编程里的惰性求值，只有在需要时计算才会发生。
forward chaining
people.eecs.berkeley.edu/~russell/cl…
Idea: fire any rule whose premises are satisfied in the KB (knowledge base), add its conclusion to the KB, until query is found
rete
A fast algorithm for the many pattern/many object pattern match problem
Key ideas:
- If-then forward chaining rules can be reorganized for efficient pattern matching.
- The RETE algorithm creates a decision tree that combines the patterns in all the rules of the knowledge base.
- Once it has been determined which patterns have been matched by facts, comparisons of variable bindings across patterns must be checked in a relational database table that ‘remembers’ what partial matches have already been tested.
- The key to its efficiency is AWP again: to do partial (incremental) matching - it remembers past test results across iterations of the rule matching loop. Only new facts are tested.
- Variable bindings are saved and reused, rather than recompjuted.
each node, except the root node, is a table representing one relation, and each row of the table records one assertions
To build a rete: • For each antecedent, create an ‘alpha’ node, aka a ‘match’ node (these are the root nodes) • Join a first antecedent and a second antecedent to create a ‘beta’ node, aka a ‘merge’ node • Join each subsequent antecedent with the previous merge node to create a new merge node • For each consequence, create a terminal node, aka execute node, via project, that carries the consequent
drool
DROOLS 是一个基于CHARLES FORGY'S的RETE算法的，易于访问企业策略、易于调整以及易于管理的开源业务规则引擎
zhuanlan.zhihu.com/p/482763434
Drools 被分为两个主要的部分：编译和运行时。编译是将规则描述文件按 ANTLR 3 语法进行解析，对语法进行正确性的检查，然后产生一种中间结构“descr”，descr 用 AST 来描述规则。目前，Drools 支持四种规则描述文件，分别是：drl 文件、 xls 文件、brl 文件和 dsl 文件，其中，常用的描述文件是 drl 文件和 xls 文件，而 xls 文件更易于维护，更直观，更为被业务人员所理解。运行时是将 AST传到 PackageBuilder，由 PackagBuilder来产生 RuleBase，它包含了一个或多个 Package 对象。
