The automation of specific mathematical tasks such as theorem proving and algebraic
manipulation have been much researched. However, there have only been a few isolated
attempts to automate the whole theory formation process. Such a process involves
forming new concepts, performing calculations, making conjectures, proving theorems
and finding counterexamples. Previous programs which perform theory formation are
limited in their functionality and their generality. We introduce the HR program
which implements a new model for theory formation. This model involves a cycle of
mathematical activity, whereby concepts are formed, conjectures about the concepts
are made and attempts to settle the conjectures are undertaken.
HR has seven general production rules for producing a new concept from old ones and
employs a best first search by building new concepts from the most interesting old
ones. To enable this, HR has various measures which estimate the interestingness of a
concept. During concept formation, HR uses empirical evidence to suggest conjectures
and employs the Otter theorem prover to attempt to prove a given conjecture. If this
fails, HR will invoke the MACE model generator to attempt to disprove the conjecture
by finding a counterexample. Information and new knowledge arising from the attempt
to settle a conjecture is used to assess the concepts involved in the conjecture, which
fuels the heuristic search and closes the cycle.
The main aim of the project has been to develop our model of theory formation and
to implement this in HR. To describe the project in the thesis, we first motivate
the problem of automated theory formation and survey the literature in this area.
We then discuss how HR invents concepts, makes and settles conjectures and how
it assesses the concepts and conjectures to facilitate a heuristic search. We present
results to evaluate HR in terms of the quality of the theories it produces and the
effectiveness of its techniques. A secondary aim of the project has been to apply HR to
mathematical discovery and we discuss how HR has successfully invented new concepts
and conjectures in number theory.