src/HOL/Tools/Sledgehammer/sledgehammer_mepo.ML
author blanchet
Fri Jul 20 22:19:46 2012 +0200 (2012-07-20 ago)
changeset 48406 b002cc16aa99
parent 48381 1b7d798460bb
child 48408 5493e67982ee
permissions -rw-r--r--
honor suggested MaSh weights
blanchet@48380
     1
(*  Title:      HOL/Tools/Sledgehammer/sledgehammer_mepo.ML
blanchet@38027
     2
    Author:     Jia Meng, Cambridge University Computer Laboratory and NICTA
blanchet@36393
     3
    Author:     Jasmin Blanchette, TU Muenchen
blanchet@39958
     4
blanchet@48380
     5
Sledgehammer's iterative relevance filter (MePo = Meng-Paulson).
wenzelm@33309
     6
*)
paulson@15452
     7
blanchet@48381
     8
signature SLEDGEHAMMER_MEPO =
wenzelm@16802
     9
sig
blanchet@46340
    10
  type stature = ATP_Problem_Generate.stature
blanchet@48296
    11
  type fact = Sledgehammer_Fact.fact
blanchet@48288
    12
  type params = Sledgehammer_Provers.params
blanchet@48288
    13
  type relevance_fudge = Sledgehammer_Provers.relevance_fudge
blanchet@40070
    14
blanchet@42646
    15
  val trace : bool Config.T
blanchet@48219
    16
  val pseudo_abs_name : string
blanchet@48219
    17
  val pseudo_skolem_prefix : string
blanchet@43351
    18
  val const_names_in_fact :
blanchet@43351
    19
    theory -> (string * typ -> term list -> bool * term list) -> term
blanchet@43351
    20
    -> string list
blanchet@48406
    21
  val mepo_suggested_facts :
blanchet@48288
    22
    Proof.context -> params -> string -> int -> relevance_fudge option
blanchet@48296
    23
    -> term list -> term -> fact list -> fact list
blanchet@48406
    24
  val weight_mepo_facts : fact list -> (fact * real) list
paulson@15347
    25
end;
paulson@15347
    26
blanchet@48381
    27
structure Sledgehammer_MePo : SLEDGEHAMMER_MEPO =
paulson@15347
    28
struct
paulson@15347
    29
blanchet@46320
    30
open ATP_Problem_Generate
blanchet@48296
    31
open Sledgehammer_Fact
blanchet@48288
    32
open Sledgehammer_Provers
blanchet@38652
    33
blanchet@42646
    34
val trace =
blanchet@48308
    35
  Attrib.setup_config_bool @{binding sledgehammer_filter_iter_trace} (K false)
blanchet@42646
    36
fun trace_msg ctxt msg = if Config.get ctxt trace then tracing (msg ()) else ()
blanchet@35826
    37
blanchet@37616
    38
val sledgehammer_prefix = "Sledgehammer" ^ Long_Name.separator
blanchet@48219
    39
val pseudo_abs_name = sledgehammer_prefix ^ "abs"
blanchet@48219
    40
val pseudo_skolem_prefix = sledgehammer_prefix ^ "sko"
blanchet@38992
    41
val theory_const_suffix = Long_Name.separator ^ " 1"
blanchet@37616
    42
blanchet@46073
    43
fun order_of_type (Type (@{type_name fun}, [T1, T2])) =
blanchet@38939
    44
    Int.max (order_of_type T1 + 1, order_of_type T2)
blanchet@38939
    45
  | order_of_type (Type (_, Ts)) = fold (Integer.max o order_of_type) Ts 0
blanchet@38939
    46
  | order_of_type _ = 0
blanchet@38939
    47
blanchet@38823
    48
(* An abstraction of Isabelle types and first-order terms *)
blanchet@38823
    49
datatype pattern = PVar | PApp of string * pattern list
blanchet@38939
    50
datatype ptype = PType of int * pattern list
blanchet@38744
    51
blanchet@38823
    52
fun string_for_pattern PVar = "_"
blanchet@38823
    53
  | string_for_pattern (PApp (s, ps)) =
blanchet@38823
    54
    if null ps then s else s ^ string_for_patterns ps
blanchet@38823
    55
and string_for_patterns ps = "(" ^ commas (map string_for_pattern ps) ^ ")"
blanchet@38939
    56
fun string_for_ptype (PType (_, ps)) = string_for_patterns ps
paulson@24287
    57
paulson@24287
    58
(*Is the second type an instance of the first one?*)
blanchet@38824
    59
fun match_pattern (PVar, _) = true
blanchet@38824
    60
  | match_pattern (PApp _, PVar) = false
blanchet@38824
    61
  | match_pattern (PApp (s, ps), PApp (t, qs)) =
blanchet@38824
    62
    s = t andalso match_patterns (ps, qs)
blanchet@38824
    63
and match_patterns (_, []) = true
blanchet@38824
    64
  | match_patterns ([], _) = false
blanchet@38824
    65
  | match_patterns (p :: ps, q :: qs) =
blanchet@38824
    66
    match_pattern (p, q) andalso match_patterns (ps, qs)
blanchet@38939
    67
fun match_ptype (PType (_, ps), PType (_, qs)) = match_patterns (ps, qs)
paulson@24287
    68
blanchet@38823
    69
(* Is there a unifiable constant? *)
blanchet@38827
    70
fun pconst_mem f consts (s, ps) =
blanchet@38939
    71
  exists (curry (match_ptype o f) ps)
blanchet@38827
    72
         (map snd (filter (curry (op =) s o fst) consts))
blanchet@38827
    73
fun pconst_hyper_mem f const_tab (s, ps) =
blanchet@38939
    74
  exists (curry (match_ptype o f) ps) (these (Symtab.lookup const_tab s))
blanchet@37505
    75
blanchet@38939
    76
fun pattern_for_type (Type (s, Ts)) = PApp (s, map pattern_for_type Ts)
blanchet@38939
    77
  | pattern_for_type (TFree (s, _)) = PApp (s, [])
blanchet@38939
    78
  | pattern_for_type (TVar _) = PVar
blanchet@38827
    79
blanchet@38744
    80
(* Pairs a constant with the list of its type instantiations. *)
blanchet@41204
    81
fun ptype thy const x =
blanchet@38939
    82
  (if const then map pattern_for_type (these (try (Sign.const_typargs thy) x))
blanchet@41204
    83
   else [])
blanchet@41204
    84
fun rich_ptype thy const (s, T) =
blanchet@41204
    85
  PType (order_of_type T, ptype thy const (s, T))
blanchet@41204
    86
fun rich_pconst thy const (s, T) = (s, rich_ptype thy const (s, T))
paulson@24287
    87
blanchet@38939
    88
fun string_for_hyper_pconst (s, ps) =
blanchet@38939
    89
  s ^ "{" ^ commas (map string_for_ptype ps) ^ "}"
paulson@24287
    90
blanchet@38823
    91
(* Add a pconstant to the table, but a [] entry means a standard
blanchet@38819
    92
   connective, which we ignore.*)
blanchet@41066
    93
fun add_pconst_to_table also_skolem (s, p) =
blanchet@48219
    94
  if (not also_skolem andalso String.isPrefix pseudo_skolem_prefix s) then I
blanchet@41066
    95
  else Symtab.map_default (s, [p]) (insert (op =) p)
blanchet@38819
    96
blanchet@47933
    97
(* Set constants tend to pull in too many irrelevant facts. We limit the damage
blanchet@47933
    98
   by treating them more or less as if they were built-in but add their
blanchet@47933
    99
   axiomatization at the end. *)
blanchet@47933
   100
val set_consts = [@{const_name Collect}, @{const_name Set.member}]
blanchet@47939
   101
val set_thms = @{thms Collect_mem_eq mem_Collect_eq Collect_cong}
blanchet@47933
   102
blanchet@42732
   103
fun add_pconsts_in_term thy is_built_in_const also_skolems pos =
blanchet@37505
   104
  let
blanchet@38819
   105
    val flip = Option.map not
blanchet@38587
   106
    (* We include free variables, as well as constants, to handle locales. For
blanchet@41205
   107
       each quantifiers that must necessarily be skolemized by the automatic
blanchet@41205
   108
       prover, we introduce a fresh constant to simulate the effect of
blanchet@41205
   109
       Skolemization. *)
blanchet@47933
   110
    fun do_const const ext_arg (x as (s, _)) ts =
blanchet@41336
   111
      let val (built_in, ts) = is_built_in_const x ts in
blanchet@47933
   112
        if member (op =) set_consts s then
blanchet@47933
   113
          fold (do_term ext_arg) ts
blanchet@47933
   114
        else
blanchet@47933
   115
          (not built_in
blanchet@47933
   116
           ? add_pconst_to_table also_skolems (rich_pconst thy const x))
blanchet@47933
   117
          #> fold (do_term false) ts
blanchet@41336
   118
      end
blanchet@42741
   119
    and do_term ext_arg t =
blanchet@38827
   120
      case strip_comb t of
blanchet@47933
   121
        (Const x, ts) => do_const true ext_arg x ts
blanchet@47933
   122
      | (Free x, ts) => do_const false ext_arg x ts
blanchet@38939
   123
      | (Abs (_, T, t'), ts) =>
blanchet@42741
   124
        ((null ts andalso not ext_arg)
blanchet@42741
   125
         (* Since lambdas on the right-hand side of equalities are usually
blanchet@47953
   126
            extensionalized later by "abs_extensionalize_term", we don't
blanchet@47953
   127
            penalize them here. *)
blanchet@48219
   128
         ? add_pconst_to_table true (pseudo_abs_name,
blanchet@48219
   129
                                     PType (order_of_type T + 1, [])))
blanchet@42735
   130
        #> fold (do_term false) (t' :: ts)
blanchet@42735
   131
      | (_, ts) => fold (do_term false) ts
blanchet@38939
   132
    fun do_quantifier will_surely_be_skolemized abs_T body_t =
blanchet@37537
   133
      do_formula pos body_t
blanchet@38747
   134
      #> (if also_skolems andalso will_surely_be_skolemized then
blanchet@48219
   135
            add_pconst_to_table true (pseudo_skolem_prefix ^ serial_string (),
blanchet@48219
   136
                                      PType (order_of_type abs_T, []))
blanchet@38587
   137
          else
blanchet@38587
   138
            I)
blanchet@42741
   139
    and do_term_or_formula ext_arg T =
blanchet@42741
   140
      if T = HOLogic.boolT then do_formula NONE else do_term ext_arg
blanchet@37537
   141
    and do_formula pos t =
blanchet@37537
   142
      case t of
blanchet@38939
   143
        Const (@{const_name all}, _) $ Abs (_, T, t') =>
blanchet@38939
   144
        do_quantifier (pos = SOME false) T t'
blanchet@37537
   145
      | @{const "==>"} $ t1 $ t2 =>
blanchet@37537
   146
        do_formula (flip pos) t1 #> do_formula pos t2
blanchet@37537
   147
      | Const (@{const_name "=="}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@42741
   148
        do_term_or_formula false T t1 #> do_term_or_formula true T t2
blanchet@37537
   149
      | @{const Trueprop} $ t1 => do_formula pos t1
blanchet@41140
   150
      | @{const False} => I
blanchet@41140
   151
      | @{const True} => I
blanchet@37537
   152
      | @{const Not} $ t1 => do_formula (flip pos) t1
blanchet@38939
   153
      | Const (@{const_name All}, _) $ Abs (_, T, t') =>
blanchet@38939
   154
        do_quantifier (pos = SOME false) T t'
blanchet@38939
   155
      | Const (@{const_name Ex}, _) $ Abs (_, T, t') =>
blanchet@38939
   156
        do_quantifier (pos = SOME true) T t'
haftmann@38795
   157
      | @{const HOL.conj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38795
   158
      | @{const HOL.disj} $ t1 $ t2 => fold (do_formula pos) [t1, t2]
haftmann@38786
   159
      | @{const HOL.implies} $ t1 $ t2 =>
blanchet@37537
   160
        do_formula (flip pos) t1 #> do_formula pos t2
haftmann@38864
   161
      | Const (@{const_name HOL.eq}, Type (_, [T, _])) $ t1 $ t2 =>
blanchet@42741
   162
        do_term_or_formula false T t1 #> do_term_or_formula true T t2
blanchet@38587
   163
      | Const (@{const_name If}, Type (_, [_, Type (_, [T, _])]))
blanchet@38587
   164
        $ t1 $ t2 $ t3 =>
blanchet@42735
   165
        do_formula NONE t1 #> fold (do_term_or_formula false T) [t2, t3]
blanchet@38939
   166
      | Const (@{const_name Ex1}, _) $ Abs (_, T, t') =>
blanchet@38939
   167
        do_quantifier (is_some pos) T t'
blanchet@38939
   168
      | Const (@{const_name Ball}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   169
        do_quantifier (pos = SOME false) T
blanchet@38939
   170
                      (HOLogic.mk_imp (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@38939
   171
      | Const (@{const_name Bex}, _) $ t1 $ Abs (_, T, t') =>
blanchet@38939
   172
        do_quantifier (pos = SOME true) T
blanchet@38939
   173
                      (HOLogic.mk_conj (incr_boundvars 1 t1 $ Bound 0, t'))
blanchet@37537
   174
      | (t0 as Const (_, @{typ bool})) $ t1 =>
blanchet@42735
   175
        do_term false t0 #> do_formula pos t1  (* theory constant *)
blanchet@42735
   176
      | _ => do_term false t
blanchet@42732
   177
  in do_formula pos end
paulson@24287
   178
blanchet@48227
   179
fun pconsts_in_fact thy is_built_in_const t =
blanchet@48227
   180
  Symtab.fold (fn (s, pss) => fold (cons o pair s) pss)
blanchet@48227
   181
              (Symtab.empty |> add_pconsts_in_term thy is_built_in_const true
blanchet@48227
   182
                                                   (SOME true) t) []
blanchet@48227
   183
blanchet@48227
   184
val const_names_in_fact = map fst ooo pconsts_in_fact
blanchet@48227
   185
blanchet@48227
   186
(* Inserts a dummy "constant" referring to the theory name, so that relevance
blanchet@48227
   187
   takes the given theory into account. *)
blanchet@41200
   188
fun theory_constify ({theory_const_rel_weight, theory_const_irrel_weight, ...}
blanchet@41200
   189
                     : relevance_fudge) thy_name t =
blanchet@40070
   190
  if exists (curry (op <) 0.0) [theory_const_rel_weight,
blanchet@40070
   191
                                theory_const_irrel_weight] then
blanchet@41200
   192
    Const (thy_name ^ theory_const_suffix, @{typ bool}) $ t
blanchet@37505
   193
  else
blanchet@41200
   194
    t
blanchet@41200
   195
blanchet@41200
   196
fun theory_const_prop_of fudge th =
blanchet@41200
   197
  theory_constify fudge (Context.theory_name (theory_of_thm th)) (prop_of th)
blanchet@37505
   198
blanchet@48227
   199
fun pair_consts_fact thy is_built_in_const fudge fact =
blanchet@48227
   200
  case fact |> snd |> theory_const_prop_of fudge
blanchet@48227
   201
            |> pconsts_in_fact thy is_built_in_const of
blanchet@48227
   202
    [] => NONE
blanchet@48227
   203
  | consts => SOME ((fact, consts), NONE)
blanchet@48227
   204
blanchet@38743
   205
(* A two-dimensional symbol table counts frequencies of constants. It's keyed
blanchet@38743
   206
   first by constant name and second by its list of type instantiations. For the
blanchet@38823
   207
   latter, we need a linear ordering on "pattern list". *)
paulson@24287
   208
blanchet@38823
   209
fun pattern_ord p =
blanchet@38743
   210
  case p of
blanchet@38744
   211
    (PVar, PVar) => EQUAL
blanchet@38823
   212
  | (PVar, PApp _) => LESS
blanchet@38823
   213
  | (PApp _, PVar) => GREATER
blanchet@38823
   214
  | (PApp q1, PApp q2) =>
blanchet@38823
   215
    prod_ord fast_string_ord (dict_ord pattern_ord) (q1, q2)
blanchet@38939
   216
fun ptype_ord (PType p, PType q) =
blanchet@38939
   217
  prod_ord (dict_ord pattern_ord) int_ord (swap p, swap q)
paulson@24287
   218
blanchet@38939
   219
structure PType_Tab = Table(type key = ptype val ord = ptype_ord)
paulson@24287
   220
blanchet@40204
   221
fun count_fact_consts thy fudge =
blanchet@37503
   222
  let
blanchet@38827
   223
    fun do_const const (s, T) ts =
blanchet@38827
   224
      (* Two-dimensional table update. Constant maps to types maps to count. *)
blanchet@41204
   225
      PType_Tab.map_default (rich_ptype thy const (s, T), 0) (Integer.add 1)
blanchet@38939
   226
      |> Symtab.map_default (s, PType_Tab.empty)
blanchet@38827
   227
      #> fold do_term ts
blanchet@38827
   228
    and do_term t =
blanchet@38827
   229
      case strip_comb t of
blanchet@38827
   230
        (Const x, ts) => do_const true x ts
blanchet@38827
   231
      | (Free x, ts) => do_const false x ts
blanchet@38827
   232
      | (Abs (_, _, t'), ts) => fold do_term (t' :: ts)
blanchet@38827
   233
      | (_, ts) => fold do_term ts
blanchet@40070
   234
  in do_term o theory_const_prop_of fudge o snd end
paulson@24287
   235
blanchet@39367
   236
fun pow_int _ 0 = 1.0
blanchet@38939
   237
  | pow_int x 1 = x
blanchet@38939
   238
  | pow_int x n = if n > 0 then x * pow_int x (n - 1) else pow_int x (n + 1) / x
blanchet@38939
   239
paulson@24287
   240
(*The frequency of a constant is the sum of those of all instances of its type.*)
blanchet@38824
   241
fun pconst_freq match const_tab (c, ps) =
blanchet@38939
   242
  PType_Tab.fold (fn (qs, m) => match (ps, qs) ? Integer.add m)
blanchet@38939
   243
                 (the (Symtab.lookup const_tab c)) 0
blanchet@38686
   244
paulson@24287
   245
blanchet@38085
   246
(* A surprising number of theorems contain only a few significant constants.
blanchet@38085
   247
   These include all induction rules, and other general theorems. *)
blanchet@37503
   248
blanchet@37503
   249
(* "log" seems best in practice. A constant function of one ignores the constant
blanchet@38938
   250
   frequencies. Rare constants give more points if they are relevant than less
blanchet@38938
   251
   rare ones. *)
blanchet@39367
   252
fun rel_weight_for _ freq = 1.0 + 2.0 / Math.ln (Real.fromInt freq + 1.0)
blanchet@38938
   253
blanchet@38938
   254
(* Irrelevant constants are treated differently. We associate lower penalties to
blanchet@38938
   255
   very rare constants and very common ones -- the former because they can't
blanchet@38938
   256
   lead to the inclusion of too many new facts, and the latter because they are
blanchet@38938
   257
   so common as to be of little interest. *)
blanchet@40070
   258
fun irrel_weight_for ({worse_irrel_freq, higher_order_irrel_weight, ...}
blanchet@40070
   259
                      : relevance_fudge) order freq =
blanchet@40070
   260
  let val (k, x) = worse_irrel_freq |> `Real.ceil in
blanchet@38939
   261
    (if freq < k then Math.ln (Real.fromInt (freq + 1)) / Math.ln x
blanchet@38939
   262
     else rel_weight_for order freq / rel_weight_for order k)
blanchet@40070
   263
    * pow_int higher_order_irrel_weight (order - 1)
blanchet@38938
   264
  end
blanchet@37503
   265
blanchet@41790
   266
fun multiplier_for_const_name local_const_multiplier s =
blanchet@41790
   267
  if String.isSubstring "." s then 1.0 else local_const_multiplier
blanchet@38821
   268
blanchet@41790
   269
(* Computes a constant's weight, as determined by its frequency. *)
blanchet@41790
   270
fun generic_pconst_weight local_const_multiplier abs_weight skolem_weight
blanchet@42735
   271
                          theory_const_weight chained_const_weight weight_for f
blanchet@42735
   272
                          const_tab chained_const_tab (c as (s, PType (m, _))) =
blanchet@48219
   273
  if s = pseudo_abs_name then
blanchet@41790
   274
    abs_weight
blanchet@48219
   275
  else if String.isPrefix pseudo_skolem_prefix s then
blanchet@41790
   276
    skolem_weight
blanchet@41790
   277
  else if String.isSuffix theory_const_suffix s then
blanchet@41790
   278
    theory_const_weight
blanchet@41790
   279
  else
blanchet@41790
   280
    multiplier_for_const_name local_const_multiplier s
blanchet@41790
   281
    * weight_for m (pconst_freq (match_ptype o f) const_tab c)
blanchet@42735
   282
    |> (if chained_const_weight < 1.0 andalso
blanchet@42735
   283
           pconst_hyper_mem I chained_const_tab c then
blanchet@42735
   284
          curry (op *) chained_const_weight
blanchet@42735
   285
        else
blanchet@42735
   286
          I)
blanchet@41790
   287
blanchet@41790
   288
fun rel_pconst_weight ({local_const_multiplier, abs_rel_weight,
blanchet@41790
   289
                        theory_const_rel_weight, ...} : relevance_fudge)
blanchet@41790
   290
                      const_tab =
blanchet@41790
   291
  generic_pconst_weight local_const_multiplier abs_rel_weight 0.0
blanchet@42735
   292
                        theory_const_rel_weight 0.0 rel_weight_for I const_tab
blanchet@42735
   293
                        Symtab.empty
blanchet@42735
   294
blanchet@41790
   295
fun irrel_pconst_weight (fudge as {local_const_multiplier, abs_irrel_weight,
blanchet@41790
   296
                                   skolem_irrel_weight,
blanchet@42735
   297
                                   theory_const_irrel_weight,
blanchet@42735
   298
                                   chained_const_irrel_weight, ...})
blanchet@42735
   299
                        const_tab chained_const_tab =
blanchet@41790
   300
  generic_pconst_weight local_const_multiplier abs_irrel_weight
blanchet@41790
   301
                        skolem_irrel_weight theory_const_irrel_weight
blanchet@42735
   302
                        chained_const_irrel_weight (irrel_weight_for fudge) swap
blanchet@42735
   303
                        const_tab chained_const_tab
paulson@24287
   304
blanchet@46340
   305
fun stature_bonus ({intro_bonus, ...} : relevance_fudge) (_, Intro) =
blanchet@46340
   306
    intro_bonus
blanchet@46340
   307
  | stature_bonus {elim_bonus, ...} (_, Elim) = elim_bonus
blanchet@46340
   308
  | stature_bonus {simp_bonus, ...} (_, Simp) = simp_bonus
blanchet@46340
   309
  | stature_bonus {local_bonus, ...} (Local, _) = local_bonus
blanchet@46340
   310
  | stature_bonus {assum_bonus, ...} (Assum, _) = assum_bonus
blanchet@46340
   311
  | stature_bonus {chained_bonus, ...} (Chained, _) = chained_bonus
blanchet@46340
   312
  | stature_bonus _ _ = 0.0
blanchet@38751
   313
blanchet@40418
   314
fun is_odd_const_name s =
blanchet@48219
   315
  s = pseudo_abs_name orelse String.isPrefix pseudo_skolem_prefix s orelse
blanchet@40418
   316
  String.isSuffix theory_const_suffix s
blanchet@40418
   317
blanchet@46340
   318
fun fact_weight fudge stature const_tab relevant_consts chained_consts
blanchet@46340
   319
                fact_consts =
blanchet@40204
   320
  case fact_consts |> List.partition (pconst_hyper_mem I relevant_consts)
blanchet@40204
   321
                   ||> filter_out (pconst_hyper_mem swap relevant_consts) of
blanchet@38827
   322
    ([], _) => 0.0
blanchet@38744
   323
  | (rel, irrel) =>
blanchet@40418
   324
    if forall (forall (is_odd_const_name o fst)) [rel, irrel] then
blanchet@40371
   325
      0.0
blanchet@40371
   326
    else
blanchet@40371
   327
      let
blanchet@42735
   328
        val irrel = irrel |> filter_out (pconst_mem swap rel)
blanchet@40371
   329
        val rel_weight =
blanchet@40371
   330
          0.0 |> fold (curry (op +) o rel_pconst_weight fudge const_tab) rel
blanchet@40371
   331
        val irrel_weight =
blanchet@46340
   332
          ~ (stature_bonus fudge stature)
blanchet@42735
   333
          |> fold (curry (op +)
blanchet@42735
   334
                   o irrel_pconst_weight fudge const_tab chained_consts) irrel
blanchet@40371
   335
        val res = rel_weight / (rel_weight + irrel_weight)
blanchet@40371
   336
      in if Real.isFinite res then res else 0.0 end
blanchet@38747
   337
blanchet@48293
   338
fun take_most_relevant ctxt max_facts remaining_max
blanchet@42728
   339
        ({max_imperfect, max_imperfect_exp, ...} : relevance_fudge)
blanchet@48296
   340
        (candidates : ((fact * (string * ptype) list) * real) list) =
blanchet@38744
   341
  let
blanchet@38747
   342
    val max_imperfect =
blanchet@40070
   343
      Real.ceil (Math.pow (max_imperfect,
blanchet@38904
   344
                    Math.pow (Real.fromInt remaining_max
blanchet@48293
   345
                              / Real.fromInt max_facts, max_imperfect_exp)))
blanchet@38747
   346
    val (perfect, imperfect) =
blanchet@38889
   347
      candidates |> sort (Real.compare o swap o pairself snd)
blanchet@38889
   348
                 |> take_prefix (fn (_, w) => w > 0.99999)
blanchet@38747
   349
    val ((accepts, more_rejects), rejects) =
blanchet@38747
   350
      chop max_imperfect imperfect |>> append perfect |>> chop remaining_max
blanchet@38744
   351
  in
blanchet@42646
   352
    trace_msg ctxt (fn () =>
wenzelm@41491
   353
        "Actually passed (" ^ string_of_int (length accepts) ^ " of " ^
wenzelm@41491
   354
        string_of_int (length candidates) ^ "): " ^
blanchet@38889
   355
        (accepts |> map (fn ((((name, _), _), _), weight) =>
blanchet@38752
   356
                            name () ^ " [" ^ Real.toString weight ^ "]")
blanchet@38745
   357
                 |> commas));
blanchet@38747
   358
    (accepts, more_rejects @ rejects)
blanchet@38744
   359
  end
paulson@24287
   360
blanchet@46340
   361
fun if_empty_replace_with_scope thy is_built_in_const facts sc tab =
blanchet@38819
   362
  if Symtab.is_empty tab then
blanchet@42732
   363
    Symtab.empty
blanchet@42732
   364
    |> fold (add_pconsts_in_term thy is_built_in_const false (SOME false))
blanchet@46340
   365
            (map_filter (fn ((_, (sc', _)), th) =>
blanchet@46340
   366
                            if sc' = sc then SOME (prop_of th) else NONE) facts)
blanchet@38819
   367
  else
blanchet@38819
   368
    tab
blanchet@38819
   369
blanchet@42702
   370
fun consider_arities is_built_in_const th =
blanchet@41158
   371
  let
blanchet@41158
   372
    fun aux _ _ NONE = NONE
blanchet@41158
   373
      | aux t args (SOME tab) =
blanchet@41158
   374
        case t of
blanchet@41158
   375
          t1 $ t2 => SOME tab |> aux t1 (t2 :: args) |> aux t2 []
blanchet@41158
   376
        | Const (x as (s, _)) =>
blanchet@41336
   377
          (if is_built_in_const x args |> fst then
blanchet@41158
   378
             SOME tab
blanchet@41158
   379
           else case Symtab.lookup tab s of
blanchet@41158
   380
             NONE => SOME (Symtab.update (s, length args) tab)
blanchet@41158
   381
           | SOME n => if n = length args then SOME tab else NONE)
blanchet@41158
   382
        | _ => SOME tab
blanchet@41158
   383
  in aux (prop_of th) [] end
blanchet@41158
   384
blanchet@44785
   385
(* FIXME: This is currently only useful for polymorphic type encodings. *)
blanchet@42702
   386
fun could_benefit_from_ext is_built_in_const facts =
blanchet@42702
   387
  fold (consider_arities is_built_in_const o snd) facts (SOME Symtab.empty)
blanchet@41158
   388
  |> is_none
blanchet@41158
   389
blanchet@43492
   390
(* High enough so that it isn't wrongly considered as very relevant (e.g., for E
blanchet@43492
   391
   weights), but low enough so that it is unlikely to be truncated away if few
blanchet@43492
   392
   facts are included. *)
blanchet@43492
   393
val special_fact_index = 75
blanchet@43492
   394
blanchet@48293
   395
fun relevance_filter ctxt thres0 decay max_facts is_built_in_const
blanchet@48292
   396
        (fudge as {threshold_divisor, ridiculous_threshold, ...}) facts hyp_ts
blanchet@48292
   397
        concl_t =
blanchet@38739
   398
  let
wenzelm@42361
   399
    val thy = Proof_Context.theory_of ctxt
blanchet@40204
   400
    val const_tab = fold (count_fact_consts thy fudge) facts Symtab.empty
blanchet@42732
   401
    val add_pconsts = add_pconsts_in_term thy is_built_in_const false o SOME
blanchet@48292
   402
    val chained_ts =
blanchet@48292
   403
      facts |> map_filter (fn ((_, (Chained, _)), th) => SOME (prop_of th)
blanchet@48292
   404
                            | _ => NONE)
blanchet@42732
   405
    val chained_const_tab = Symtab.empty |> fold (add_pconsts true) chained_ts
blanchet@38819
   406
    val goal_const_tab =
blanchet@42732
   407
      Symtab.empty |> fold (add_pconsts true) hyp_ts
blanchet@42732
   408
                   |> add_pconsts false concl_t
blanchet@42732
   409
      |> (fn tab => if Symtab.is_empty tab then chained_const_tab else tab)
blanchet@46340
   410
      |> fold (if_empty_replace_with_scope thy is_built_in_const facts)
blanchet@38993
   411
              [Chained, Assum, Local]
blanchet@48288
   412
    fun iter j remaining_max thres rel_const_tab hopeless hopeful =
blanchet@38739
   413
      let
blanchet@40191
   414
        fun relevant [] _ [] =
blanchet@38747
   415
            (* Nothing has been added this iteration. *)
blanchet@48288
   416
            if j = 0 andalso thres >= ridiculous_threshold then
blanchet@38747
   417
              (* First iteration? Try again. *)
blanchet@48293
   418
              iter 0 max_facts (thres / threshold_divisor) rel_const_tab
blanchet@38747
   419
                   hopeless hopeful
blanchet@38744
   420
            else
blanchet@40191
   421
              []
blanchet@38889
   422
          | relevant candidates rejects [] =
blanchet@38739
   423
            let
blanchet@38747
   424
              val (accepts, more_rejects) =
blanchet@48293
   425
                take_most_relevant ctxt max_facts remaining_max fudge candidates
blanchet@38739
   426
              val rel_const_tab' =
blanchet@38745
   427
                rel_const_tab
blanchet@41066
   428
                |> fold (add_pconst_to_table false) (maps (snd o fst) accepts)
blanchet@38744
   429
              fun is_dirty (c, _) =
blanchet@38744
   430
                Symtab.lookup rel_const_tab' c <> Symtab.lookup rel_const_tab c
blanchet@38745
   431
              val (hopeful_rejects, hopeless_rejects) =
blanchet@38745
   432
                 (rejects @ hopeless, ([], []))
blanchet@38745
   433
                 |-> fold (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   434
                              if exists is_dirty consts then
blanchet@38745
   435
                                apfst (cons (ax, NONE))
blanchet@38745
   436
                              else
blanchet@38745
   437
                                apsnd (cons (ax, old_weight)))
blanchet@38745
   438
                 |>> append (more_rejects
blanchet@38745
   439
                             |> map (fn (ax as (_, consts), old_weight) =>
blanchet@38745
   440
                                        (ax, if exists is_dirty consts then NONE
blanchet@38745
   441
                                             else SOME old_weight)))
blanchet@48288
   442
              val thres =
blanchet@48288
   443
                1.0 - (1.0 - thres)
blanchet@38822
   444
                      * Math.pow (decay, Real.fromInt (length accepts))
blanchet@38747
   445
              val remaining_max = remaining_max - length accepts
blanchet@38739
   446
            in
blanchet@42646
   447
              trace_msg ctxt (fn () => "New or updated constants: " ^
blanchet@38744
   448
                  commas (rel_const_tab' |> Symtab.dest
blanchet@38822
   449
                          |> subtract (op =) (rel_const_tab |> Symtab.dest)
blanchet@38827
   450
                          |> map string_for_hyper_pconst));
blanchet@38745
   451
              map (fst o fst) accepts @
blanchet@38747
   452
              (if remaining_max = 0 then
blanchet@40191
   453
                 []
blanchet@38745
   454
               else
blanchet@48288
   455
                 iter (j + 1) remaining_max thres rel_const_tab'
blanchet@38747
   456
                      hopeless_rejects hopeful_rejects)
blanchet@38739
   457
            end
blanchet@38889
   458
          | relevant candidates rejects
blanchet@46340
   459
                     (((ax as (((_, stature), _), fact_consts)), cached_weight)
blanchet@38747
   460
                      :: hopeful) =
blanchet@38739
   461
            let
blanchet@38739
   462
              val weight =
blanchet@38739
   463
                case cached_weight of
blanchet@38739
   464
                  SOME w => w
blanchet@46340
   465
                | NONE => fact_weight fudge stature const_tab rel_const_tab
blanchet@42732
   466
                                      chained_const_tab fact_consts
blanchet@38739
   467
            in
blanchet@48288
   468
              if weight >= thres then
blanchet@38889
   469
                relevant ((ax, weight) :: candidates) rejects hopeful
blanchet@38739
   470
              else
blanchet@38889
   471
                relevant candidates ((ax, weight) :: rejects) hopeful
blanchet@38739
   472
            end
blanchet@38739
   473
        in
blanchet@42646
   474
          trace_msg ctxt (fn () =>
blanchet@38744
   475
              "ITERATION " ^ string_of_int j ^ ": current threshold: " ^
blanchet@48288
   476
              Real.toString thres ^ ", constants: " ^
blanchet@38744
   477
              commas (rel_const_tab |> Symtab.dest
blanchet@38744
   478
                      |> filter (curry (op <>) [] o snd)
blanchet@38827
   479
                      |> map string_for_hyper_pconst));
blanchet@38889
   480
          relevant [] [] hopeful
blanchet@38739
   481
        end
blanchet@47933
   482
    fun uses_const s t =
blanchet@47933
   483
      fold_aterms (curry (fn (Const (s', _), false) => s' = s | (_, b) => b)) t
blanchet@47933
   484
                  false
blanchet@47933
   485
    fun uses_const_anywhere accepts s =
blanchet@47933
   486
      exists (uses_const s o prop_of o snd) accepts orelse
blanchet@47933
   487
      exists (uses_const s) (concl_t :: hyp_ts)
blanchet@47933
   488
    fun add_set_const_thms accepts =
blanchet@47933
   489
      exists (uses_const_anywhere accepts) set_consts ? append set_thms
blanchet@43492
   490
    fun insert_into_facts accepts [] = accepts
blanchet@43492
   491
      | insert_into_facts accepts ths =
blanchet@43492
   492
        let
blanchet@43492
   493
          val add = facts |> filter (member Thm.eq_thm_prop ths o snd)
blanchet@43492
   494
          val (bef, after) =
blanchet@43492
   495
            accepts |> filter_out (member Thm.eq_thm_prop ths o snd)
blanchet@48293
   496
                    |> take (max_facts - length add)
blanchet@43492
   497
                    |> chop special_fact_index
blanchet@43492
   498
        in bef @ add @ after end
blanchet@43492
   499
    fun insert_special_facts accepts =
blanchet@47939
   500
       (* FIXME: get rid of "ext" here once it is treated as a helper *)
blanchet@43066
   501
       [] |> could_benefit_from_ext is_built_in_const accepts ? cons @{thm ext}
blanchet@47933
   502
          |> add_set_const_thms accepts
blanchet@43492
   503
          |> insert_into_facts accepts
blanchet@38739
   504
  in
blanchet@40369
   505
    facts |> map_filter (pair_consts_fact thy is_built_in_const fudge)
blanchet@48293
   506
          |> iter 0 max_facts thres0 goal_const_tab []
blanchet@43492
   507
          |> insert_special_facts
blanchet@42646
   508
          |> tap (fn accepts => trace_msg ctxt (fn () =>
wenzelm@41491
   509
                      "Total relevant: " ^ string_of_int (length accepts)))
blanchet@38739
   510
  end
paulson@24287
   511
blanchet@48406
   512
fun mepo_suggested_facts ctxt
blanchet@48293
   513
        ({fact_thresholds = (thres0, thres1), ...} : params) prover
blanchet@48293
   514
        max_facts fudge hyp_ts concl_t facts =
blanchet@37538
   515
  let
wenzelm@42361
   516
    val thy = Proof_Context.theory_of ctxt
blanchet@48288
   517
    val is_built_in_const =
blanchet@48288
   518
      Sledgehammer_Provers.is_built_in_const_for_prover ctxt prover
blanchet@48288
   519
    val fudge =
blanchet@48288
   520
      case fudge of
blanchet@48288
   521
        SOME fudge => fudge
blanchet@48288
   522
      | NONE => Sledgehammer_Provers.relevance_fudge_for_prover ctxt prover
blanchet@48288
   523
    val decay = Math.pow ((1.0 - thres1) / (1.0 - thres0),
blanchet@48293
   524
                          1.0 / Real.fromInt (max_facts + 1))
blanchet@37538
   525
  in
blanchet@42646
   526
    trace_msg ctxt (fn () => "Considering " ^ string_of_int (length facts) ^
blanchet@42646
   527
                             " facts");
blanchet@48288
   528
    (if thres1 < 0.0 then
blanchet@40204
   529
       facts
blanchet@48288
   530
     else if thres0 > 1.0 orelse thres0 > thres1 then
blanchet@38739
   531
       []
blanchet@38739
   532
     else
blanchet@48293
   533
       relevance_filter ctxt thres0 decay max_facts is_built_in_const fudge
blanchet@48293
   534
           facts hyp_ts
blanchet@44625
   535
           (concl_t |> theory_constify fudge (Context.theory_name thy)))
blanchet@37538
   536
  end
immler@30536
   537
blanchet@48406
   538
(* Ad hoc score function roughly based on Blanchette's Ringberg 2011 data. *)
blanchet@48406
   539
fun weight_of_fact rank =
blanchet@48406
   540
  Math.pow (1.5, 15.5 - 0.05 * Real.fromInt rank) + 15.0
blanchet@48406
   541
blanchet@48406
   542
fun weight_mepo_facts facts =
blanchet@48406
   543
  facts ~~ map weight_of_fact (0 upto length facts - 1)
blanchet@48406
   544
paulson@15347
   545
end;