Forward Chaining vs Backward Chaining

Question

What is one good for that the other's not in practice? I understand the theory of what they do, but what are their limitations and capabilities in practical use? I'm considering Drools vs a java prolog for a new AI project, but open to other suggestions. What are some popular approaches for inferencing on a complicated relational data set or alternatives?

Answer 1

Backward chaining (a la Prolog) is more like finding what initial conditions form a path to your goal. At a very basic level it is a backward search from your goal to find conditions that will fulfil it.

Backward chaining is used for interrogative applications (finding items that fulfil certain criteria) - one commercial example of a backward chaining application might be finding which insurance policies are covered by a particular reinsurance contract.

Forward chaining (a la CLIPS) matches conditions and then generates inferences from those conditions. These conditions can in turn match other rules. Basically, this takes a set of initial conditions and then draws all inferences it can from those conditions.

The inferences (if asserted) can also be actions or events that can trigger external actions. This is useful in event driven systems, as the rule sets can be configured to (for example) initiate a workflow or some other action. This type of rule engine is the most commonly used in commercial applications.

Event driven systems are a common application of forward chaining rule engines. One example of a forward chaining application might be a telecoms plan provisioning engine (typically used for administering mobile phone plans). Entering a particular user with a particular plan will trigger a range of items to be set up in various phone switches, billing systems, financials, CRM systems etc.

Answer 2

Concerned's answer is very good. When asked to boil the difference down to a sound bite, I usually say something like:

Lots of Output Hypotheses + Lots of Data Up Front => Use Forward Chaining

Fewer Output Hypotheses + Must Query for Data => Use Backward Chaining

But it's just a rule of thumb, not a commandment.