Next: Bibliography Up: COMP9417 Project Shapiro's Model Previous: Examples of MIS Learning

Subsections



Conclusion

Having presented Shapiro's Model Inference System, we can now address some of the questions posed in the introduction:

These last three issues indicate that we must go beyond pure classical first-order logic to find a learning algorithm usable on practical, non-trivial domains. The most obvious recourse is the same as that for the physical sciences: apply statistics. See [MR99] for some recent work on this.


Multiple Predicate Learning

Some concepts cannot be learnt with a single set of clauses that all have the same head symbol (that is, define a single predicate). An example of this is anywhere where two or more levels of recursion (analogous to nested or sequenced loops) are required to express the relation, such as in successor-arithmetic multiplication, and a list-of-lists to lists flattening transformation.

Hence, it is necessary that auxilliary predicates can be introduced. If the learning algorithm can do this automatically, it is said to do predicate invention.

MIS will not create such theoretical terms on it's own initiative, although it can do multi-predicate learning - which is to learn several predicates as part of the one learning task. Additionally, MIS can either learn these predicates concurrently, or learn one and then use it as background knowledge for a later one. In either case, the learning task must be structured, as the oracle must be able to answer queries on every predicate in the target theory.


MIS Against FOIL

FOIL [QCJ93] is a well known ILP system that has quite distinct properties from MIS.


next up previous
Next: Bibliography Up: COMP9417 Project Shapiro's Model Previous: Examples of MIS Learning
Peter Gammie
2002-03-01