Notes on supplementary materials:

- supp.pdf contains a short description of the APIs of the environment including the visualization of proof search, an analysis of the argument space and example proofs found by TacticZero.

- directory ./proof_search_trees contains 5 automatically generated HTML files representing 5 different proofs found by TacticZero. Each graph is essentially a terminal state of the MDP (defined in the paper) if the edges are ignored. 

- theorems.pdf lists the 1342 theorems we used for our experiments.  

Method:

For each theorem in the core library of HOL, we test whether it is provable by the selected tactics in table 1 of the main submission.

This is done by applying random tactics from table 1 for a fixed number of steps, if the theorem is proved, it is then put into the dataset. If not, we proceed with next theorem in the library. We iterate this process for the entire core library for a month, and collect all the theorems proved this way. This approach enables us to discover provable theorems whose human proofs appear to be too complicated to be proved by the selected tactics. Most of the discovered theorems are also non-trivial, as suggested by figure 5 in the main submission.
