Structure PmatchHeuristics

signature PmatchHeuristics =
sig

   type term = Term.term
   type thm = Thm.thm
   type thry = {Thy : string, Tyop : string} ->
               {case_const : term, constructors : term list} option
   type pmatch_heuristic = {skip_rows : bool, (* skip splitting for redundant rows? *)
                            collapse_cases : bool, (* collapse cases that lead to the same result ? *)
                            (* given a list of rows of patterns, which column to split on? *)
                            col_fun : thry -> term list list -> int }

   (* some predefined heuristics *)
   val pheu_classic : pmatch_heuristic (* HOL 4's old heuristic *)
   val pheu_first_row : pmatch_heuristic
   val pheu_constr_prefix : pmatch_heuristic
   val pheu_qba : pmatch_heuristic  (* the recommended one *)
   val pheu_cqba : pmatch_heuristic
   val pheu_first_col : pmatch_heuristic
   val pheu_last_col : pmatch_heuristic

   (* A manual heuristic. For each decision, it prints the columns and
      takes the column-number from the provided list of explicit
      choices. If the list is too short, the first column is
      chosen. One should run this heuristic first with an empty list
      as argument. Then look at the choices and add 0s for all the
      initial choices you liked to be 0 an then a different choice you
      prefer. Rerun again (because the following choices will change)
      and iterate. This provided very fine control, but is tedious. *)
   val pheu_manual : int list -> pmatch_heuristic

   (* A heuristic based on column ranks. Given a pattern match matrix like

      p_11 ... p_1n
      ...
      p_m1 --- p_mn

      and a list of ranking functions prheuL = [r_1, ... r_j]. The
      heuristic pheu_rank applies all ranking functions to all columns.
      Let's denote the result of "r_i current_thyr [p_k1, ... pkm]" with
      c_ik. It then picks column i such that [c_1i, ... c_ji] is maximal
      accroding to the lexicographic ordering of integers.
   *)
   val pheu_rank : (thry -> term list -> int) list -> pmatch_heuristic

   (* some ranking functions *)
   val prheu_first_row : thry -> term list -> int
   val prheu_constr_prefix : thry -> term list -> int
   val prheu_small_branching_factor : thry -> term list -> int
   val prheu_arity : thry -> term list -> int

   (* A comparison for the results of heuristic application
      (list of pattern lists, resulting decision tree) *)
   type pmatch_heuristic_res_compare = (term list list * term) Lib.cmp
   val pmatch_heuristic_cases_cmp : pmatch_heuristic_res_compare (* few cases are good *)
   val pmatch_heuristic_size_cmp : pmatch_heuristic_res_compare (* small terms are good *)

   (* Using such comparisons, we can supply multiple heuristics and choose the best results.
      For technical reasons, this function might be stateful and therefore get's unit arguments.

      The usage of a heu_fun by the Pmatch library is as follows. Heu_fun initialises the functions and returns a
      compare function and a function heu_fun' which provides heuristics. As long as
      heu_fun' () provides fresh heuristics these are tried. Then the best result of all these
      heuristics according to the compare function is choosen. *)
   type pmatch_heuristic_fun = unit -> pmatch_heuristic_res_compare * (unit -> pmatch_heuristic option)

   val default_heuristic_fun : pmatch_heuristic_fun
   val classic_heuristic_fun : pmatch_heuristic_fun

   (* An exhaustive heuristic_fun. It tries all possibilities and very quickly blows up!
      Only usable for very small examples! *)
   val exhaustive_heuristic_fun : pmatch_heuristic_res_compare -> pmatch_heuristic_fun


   (* custom pmatch_heuristic_fun can be easiest constructed by an explicit list of heuristics and
      a compare function *)
   val pmatch_heuristic_list : pmatch_heuristic_res_compare -> pmatch_heuristic list -> pmatch_heuristic_fun

   (* A list of useful heuristics to be used with pmatch_heuristic_list *)
   val default_heuristic_list : pmatch_heuristic list


   (* The pmatch_heuristic_fun to be used by default and various functions to set it *)
   val pmatch_heuristic : pmatch_heuristic_fun ref

   val set_heuristic : pmatch_heuristic -> unit
   val set_heuristic_fun : pmatch_heuristic_fun -> unit
   val set_heuristic_list_size : pmatch_heuristic list -> unit
   val set_heuristic_list_cases : pmatch_heuristic list -> unit
   val set_default_heuristic : unit -> unit
   val set_classic_heuristic : unit -> unit

   val with_classic_heuristic : ('a -> 'b) -> ('a -> 'b)
   val with_heuristic : pmatch_heuristic -> ('a -> 'b) -> ('a -> 'b)
   val with_manual_heuristic : int list -> ('a -> 'b) -> ('a -> 'b)
   val is_classic : unit -> bool
end

HOL 4, Kananaskis-11