How the Inference Engine Works › Backward Chaining

Backward Chaining

Posting is the same as in forward chaining, and all posted rules start in the Ready state. Since backward chaining is goal-driven, however, the engine processes rules a little differently-it starts by creating a goal list. Initially, the goal list contains only the main goal as specified by the chaining statement.

The engine visits the first rule in priority order and ascertains if the rule action contains an assignment to the goal attribute. If not, the engine moves on to the next rule in priority order.

If so, the engine tries to evaluate the premise:

• A TRUE premise causes the rule to fire. The rule's state changes to Fired.
• A FALSE premise causes the rule to fail. The rule's state changes to Failed.
• If the premise contains one or more UNKNOWN attributes, they are added to the goal list as subgoals. The rule's state changes to Pended. The engine then goes back to the start of the priority order and visits the first Ready rule.

The engine proceeds through the Ready rules in priority order. With each rule, it checks to see if the action contains assignments to any of the attributes in the goal list. If so, the engine tries to evaluate the rule premise. If not, the rule is passed over.

The engine starts back at the top of the priority order whenever the following happens:

• A rule is fired, and its action makes an assignment to a subgoal. All pended rules that contain the subgoal change to the Ready state.
• A rule is pended. The UNKNOWN attribute is added to the goal list.

As in forward chaining, the engine continues until the main goal is assigned, no Ready rules remain, or you issue the stopchain command. Backward chaining requires, however, that you specify a goal attribute.