src/HOL/Tools/Sledgehammer/sledgehammer_mash.ML
author blanchet
Fri, 20 Jul 2012 22:19:46 +0200
changeset 49398 df75b2d7e26a
parent 49397 641af72b0059
child 49399 83dc102041e6
permissions -rw-r--r--
learn command in MaSh
     1 (*  Title:      HOL/Tools/Sledgehammer/sledgehammer_mash.ML
     2     Author:     Jasmin Blanchette, TU Muenchen
     3 
     4 Sledgehammer's machine-learning-based relevance filter (MaSh).
     5 *)
     6 
     7 signature SLEDGEHAMMER_MASH =
     8 sig
     9   type status = ATP_Problem_Generate.status
    10   type stature = ATP_Problem_Generate.stature
    11   type fact = Sledgehammer_Fact.fact
    12   type fact_override = Sledgehammer_Fact.fact_override
    13   type params = Sledgehammer_Provers.params
    14   type relevance_fudge = Sledgehammer_Provers.relevance_fudge
    15   type prover_result = Sledgehammer_Provers.prover_result
    16 
    17   val trace : bool Config.T
    18   val MaShN : string
    19   val mepoN : string
    20   val mashN : string
    21   val meshN : string
    22   val fact_filters : string list
    23   val escape_meta : string -> string
    24   val escape_metas : string list -> string
    25   val unescape_meta : string -> string
    26   val unescape_metas : string -> string list
    27   val extract_query : string -> string * string list
    28   val nickname_of : thm -> string
    29   val suggested_facts : string list -> ('a * thm) list -> ('a * thm) list
    30   val mesh_facts :
    31     int -> (('a * thm) list * ('a * thm) list) list -> ('a * thm) list
    32   val is_likely_tautology_or_too_meta : thm -> bool
    33   val theory_ord : theory * theory -> order
    34   val thm_ord : thm * thm -> order
    35   val features_of :
    36     Proof.context -> string -> theory -> status -> term list -> string list
    37   val isabelle_dependencies_of : unit Symtab.table -> thm -> string list
    38   val goal_of_thm : theory -> thm -> thm
    39   val run_prover_for_mash :
    40     Proof.context -> params -> string -> fact list -> thm -> prover_result
    41   val mash_CLEAR : Proof.context -> unit
    42   val mash_INIT :
    43     Proof.context -> bool
    44     -> (string * string list * string list * string list) list -> unit
    45   val mash_ADD :
    46     Proof.context -> bool
    47     -> (string * string list * string list * string list) list -> unit
    48   val mash_QUERY :
    49     Proof.context -> bool -> int -> string list * string list -> string list
    50   val mash_unlearn : Proof.context -> unit
    51   val mash_could_suggest_facts : unit -> bool
    52   val mash_can_suggest_facts : Proof.context -> bool
    53   val mash_suggest_facts :
    54     Proof.context -> params -> string -> int -> term list -> term
    55     -> ('a * thm) list -> ('a * thm) list * ('a * thm) list
    56   val mash_learn_proof :
    57     Proof.context -> params -> term -> ('a * thm) list -> thm list -> unit
    58   val mash_learn_thy :
    59     Proof.context -> params -> theory -> Time.time -> fact list -> string
    60   val mash_learn : Proof.context -> params -> unit
    61   val relevant_facts :
    62     Proof.context -> params -> string -> int -> fact_override -> term list
    63     -> term -> fact list -> fact list
    64   val kill_learners : unit -> unit
    65   val running_learners : unit -> unit
    66 end;
    67 
    68 structure Sledgehammer_MaSh : SLEDGEHAMMER_MASH =
    69 struct
    70 
    71 open ATP_Util
    72 open ATP_Problem_Generate
    73 open Sledgehammer_Util
    74 open Sledgehammer_Fact
    75 open Sledgehammer_Provers
    76 open Sledgehammer_Minimize
    77 open Sledgehammer_MePo
    78 
    79 val trace =
    80   Attrib.setup_config_bool @{binding sledgehammer_mash_trace} (K false)
    81 fun trace_msg ctxt msg = if Config.get ctxt trace then tracing (msg ()) else ()
    82 
    83 val MaShN = "MaSh"
    84 
    85 val mepoN = "mepo"
    86 val mashN = "mash"
    87 val meshN = "mesh"
    88 
    89 val fact_filters = [meshN, mepoN, mashN]
    90 
    91 fun mash_home () = getenv "MASH_HOME"
    92 fun mash_state_dir () =
    93   getenv "ISABELLE_HOME_USER" ^ "/mash"
    94   |> tap (Isabelle_System.mkdir o Path.explode)
    95 fun mash_state_path () = mash_state_dir () ^ "/state" |> Path.explode
    96 
    97 
    98 (*** Isabelle helpers ***)
    99 
   100 fun meta_char c =
   101   if Char.isAlphaNum c orelse c = #"_" orelse c = #"." orelse c = #"(" orelse
   102      c = #")" orelse c = #"," then
   103     String.str c
   104   else
   105     (* fixed width, in case more digits follow *)
   106     "\\" ^ stringN_of_int 3 (Char.ord c)
   107 
   108 fun unmeta_chars accum [] = String.implode (rev accum)
   109   | unmeta_chars accum (#"\\" :: d1 :: d2 :: d3 :: cs) =
   110     (case Int.fromString (String.implode [d1, d2, d3]) of
   111        SOME n => unmeta_chars (Char.chr n :: accum) cs
   112      | NONE => "" (* error *))
   113   | unmeta_chars _ (#"\\" :: _) = "" (* error *)
   114   | unmeta_chars accum (c :: cs) = unmeta_chars (c :: accum) cs
   115 
   116 val escape_meta = String.translate meta_char
   117 val escape_metas = map escape_meta #> space_implode " "
   118 val unescape_meta = String.explode #> unmeta_chars []
   119 val unescape_metas =
   120   space_explode " " #> filter_out (curry (op =) "") #> map unescape_meta
   121 
   122 fun extract_query line =
   123   case space_explode ":" line of
   124     [goal_name, suggs] => (unescape_meta goal_name, unescape_metas suggs)
   125   | _ => ("", [])
   126 
   127 fun parent_of_local_thm th =
   128   let
   129     val thy = th |> Thm.theory_of_thm
   130     val facts = thy |> Global_Theory.facts_of
   131     val space = facts |> Facts.space_of
   132     fun id_of s = #id (Name_Space.the_entry space s)
   133     fun max_id (s', _) (s, id) =
   134       let val id' = id_of s' in if id > id' then (s, id) else (s', id') end
   135   in ("", ~1) |> Facts.fold_static max_id facts |> fst end
   136 
   137 val local_prefix = "local" ^ Long_Name.separator
   138 
   139 fun nickname_of th =
   140   let val hint = Thm.get_name_hint th in
   141     (* FIXME: There must be a better way to detect local facts. *)
   142     case try (unprefix local_prefix) hint of
   143       SOME suff =>
   144       parent_of_local_thm th ^ Long_Name.separator ^ Long_Name.separator ^ suff
   145     | NONE => hint
   146   end
   147 
   148 fun suggested_facts suggs facts =
   149   let
   150     fun add_fact (fact as (_, th)) = Symtab.default (nickname_of th, fact)
   151     val tab = Symtab.empty |> fold add_fact facts
   152   in map_filter (Symtab.lookup tab) suggs end
   153 
   154 (* Ad hoc score function roughly based on Blanchette's Ringberg 2011 data. *)
   155 fun score x = Math.pow (1.5, 15.5 - 0.05 * Real.fromInt x) + 15.0
   156 
   157 fun sum_sq_avg [] = 0
   158   | sum_sq_avg xs =
   159     Real.ceil (100000.0 * fold (curry (op +) o score) xs 0.0) div length xs
   160 
   161 fun mesh_facts max_facts [(selected, unknown)] =
   162     take max_facts selected @ take (max_facts - length selected) unknown
   163   | mesh_facts max_facts mess =
   164     let
   165       val mess = mess |> map (apfst (`length))
   166       val fact_eq = Thm.eq_thm o pairself snd
   167       fun score_in fact ((sel_len, sels), unks) =
   168         case find_index (curry fact_eq fact) sels of
   169           ~1 => (case find_index (curry fact_eq fact) unks of
   170                    ~1 => SOME sel_len
   171                  | _ => NONE)
   172         | j => SOME j
   173       fun score_of fact = mess |> map_filter (score_in fact) |> sum_sq_avg
   174       val facts = fold (union fact_eq o take max_facts o snd o fst) mess []
   175     in
   176       facts |> map (`score_of) |> sort (int_ord o swap o pairself fst)
   177             |> map snd |> take max_facts
   178     end
   179 
   180 val thy_feature_prefix = "y_"
   181 
   182 val thy_feature_name_of = prefix thy_feature_prefix
   183 val const_name_of = prefix const_prefix
   184 val type_name_of = prefix type_const_prefix
   185 val class_name_of = prefix class_prefix
   186 
   187 fun is_likely_tautology_or_too_meta th =
   188   let
   189     val is_boring_const = member (op =) atp_widely_irrelevant_consts
   190     fun is_boring_bool t =
   191       not (exists_Const (not o is_boring_const o fst) t) orelse
   192       exists_type (exists_subtype (curry (op =) @{typ prop})) t
   193     fun is_boring_prop (@{const Trueprop} $ t) = is_boring_bool t
   194       | is_boring_prop (@{const "==>"} $ t $ u) =
   195         is_boring_prop t andalso is_boring_prop u
   196       | is_boring_prop (Const (@{const_name all}, _) $ (Abs (_, _, t)) $ u) =
   197         is_boring_prop t andalso is_boring_prop u
   198       | is_boring_prop (Const (@{const_name "=="}, _) $ t $ u) =
   199         is_boring_bool t andalso is_boring_bool u
   200       | is_boring_prop _ = true
   201   in
   202     is_boring_prop (prop_of th) andalso not (Thm.eq_thm_prop (@{thm ext}, th))
   203   end
   204 
   205 fun theory_ord p =
   206   if Theory.eq_thy p then
   207     EQUAL
   208   else if Theory.subthy p then
   209     LESS
   210   else if Theory.subthy (swap p) then
   211     GREATER
   212   else case int_ord (pairself (length o Theory.ancestors_of) p) of
   213     EQUAL => string_ord (pairself Context.theory_name p)
   214   | order => order
   215 
   216 val thm_ord = theory_ord o pairself theory_of_thm
   217 
   218 val bad_types = [@{type_name prop}, @{type_name bool}, @{type_name fun}]
   219 
   220 fun interesting_terms_types_and_classes ctxt prover term_max_depth
   221                                         type_max_depth ts =
   222   let
   223     fun is_bad_const (x as (s, _)) args =
   224       member (op =) atp_logical_consts s orelse
   225       fst (is_built_in_const_for_prover ctxt prover x args)
   226     fun add_classes @{sort type} = I
   227       | add_classes S = union (op =) (map class_name_of S)
   228     fun do_add_type (Type (s, Ts)) =
   229         (not (member (op =) bad_types s) ? insert (op =) (type_name_of s))
   230         #> fold do_add_type Ts
   231       | do_add_type (TFree (_, S)) = add_classes S
   232       | do_add_type (TVar (_, S)) = add_classes S
   233     fun add_type T = type_max_depth >= 0 ? do_add_type T
   234     fun mk_app s args =
   235       if member (op <>) args "" then s ^ "(" ^ space_implode "," args ^ ")"
   236       else s
   237     fun patternify ~1 _ = ""
   238       | patternify depth t =
   239         case strip_comb t of
   240           (Const (s, _), args) =>
   241           mk_app (const_name_of s) (map (patternify (depth - 1)) args)
   242         | _ => ""
   243     fun add_term_patterns ~1 _ = I
   244       | add_term_patterns depth t =
   245         insert (op =) (patternify depth t)
   246         #> add_term_patterns (depth - 1) t
   247     val add_term = add_term_patterns term_max_depth
   248     fun add_patterns t =
   249       let val (head, args) = strip_comb t in
   250         (case head of
   251            Const (x as (_, T)) =>
   252            not (is_bad_const x args) ? (add_term t #> add_type T)
   253          | Free (_, T) => add_type T
   254          | Var (_, T) => add_type T
   255          | Abs (_, T, body) => add_type T #> add_patterns body
   256          | _ => I)
   257         #> fold add_patterns args
   258       end
   259   in [] |> fold add_patterns ts end
   260 
   261 fun is_exists (s, _) = (s = @{const_name Ex} orelse s = @{const_name Ex1})
   262 
   263 val term_max_depth = 1
   264 val type_max_depth = 1
   265 
   266 (* TODO: Generate type classes for types? *)
   267 fun features_of ctxt prover thy status ts =
   268   thy_feature_name_of (Context.theory_name thy) ::
   269   interesting_terms_types_and_classes ctxt prover term_max_depth type_max_depth
   270                                       ts
   271   |> forall is_lambda_free ts ? cons "no_lams"
   272   |> forall (not o exists_Const is_exists) ts ? cons "no_skos"
   273   |> (case status of
   274         General => I
   275       | Induction => cons "induction"
   276       | Intro => cons "intro"
   277       | Inductive => cons "inductive"
   278       | Elim => cons "elim"
   279       | Simp => cons "simp"
   280       | Def => cons "def")
   281 
   282 fun isabelle_dependencies_of all_facts = thms_in_proof (SOME all_facts)
   283 
   284 val freezeT = Type.legacy_freeze_type
   285 
   286 fun freeze (t $ u) = freeze t $ freeze u
   287   | freeze (Abs (s, T, t)) = Abs (s, freezeT T, freeze t)
   288   | freeze (Var ((s, _), T)) = Free (s, freezeT T)
   289   | freeze (Const (s, T)) = Const (s, freezeT T)
   290   | freeze (Free (s, T)) = Free (s, freezeT T)
   291   | freeze t = t
   292 
   293 fun goal_of_thm thy = prop_of #> freeze #> cterm_of thy #> Goal.init
   294 
   295 fun run_prover_for_mash ctxt params prover facts goal =
   296   let
   297     val problem =
   298       {state = Proof.init ctxt, goal = goal, subgoal = 1, subgoal_count = 1,
   299        facts = facts |> map (apfst (apfst (fn name => name ())))
   300                      |> map Untranslated_Fact}
   301     val prover = get_minimizing_prover ctxt Normal (K ()) prover
   302   in prover params (K (K (K ""))) problem end
   303 
   304 
   305 (*** Low-level communication with MaSh ***)
   306 
   307 fun write_file (xs, f) file =
   308   let val path = Path.explode file in
   309     File.write path "";
   310     xs |> chunk_list 500
   311        |> List.app (File.append path o space_implode "" o map f)
   312   end
   313 
   314 fun mash_info overlord =
   315   if overlord then (getenv "ISABELLE_HOME_USER", "")
   316   else (getenv "ISABELLE_TMP", serial_string ())
   317 
   318 fun and_rm_files overlord flags files =
   319   if overlord then
   320     ""
   321   else
   322     " && rm -f" ^ flags ^ " -- " ^
   323     space_implode " " (map File.shell_quote files)
   324 
   325 fun run_mash ctxt overlord (temp_dir, serial) async core =
   326   let
   327     val log_file = temp_dir ^ "/mash_log" ^ serial
   328     val err_file = temp_dir ^ "/mash_err" ^ serial
   329     val command =
   330       "(" ^ mash_home () ^ "/mash --quiet --outputDir " ^ mash_state_dir () ^
   331       " --log " ^ log_file ^ " " ^ core ^ ") 2>&1 > " ^ err_file ^
   332       and_rm_files overlord "" [log_file, err_file] ^
   333       (if async then " &" else "")
   334   in
   335     trace_msg ctxt (fn () =>
   336         (if async then "Launching " else "Running ") ^ command);
   337     write_file ([], K "") log_file;
   338     write_file ([], K "") err_file;
   339     Isabelle_System.bash command;
   340     if not async then trace_msg ctxt (K "Done") else ()
   341   end
   342 
   343 (* TODO: Eliminate code once "mash.py" handles sequences of ADD commands as fast
   344    as a single INIT. *)
   345 fun run_mash_init ctxt overlord write_access write_feats write_deps =
   346   let
   347     val info as (temp_dir, serial) = mash_info overlord
   348     val in_dir = temp_dir ^ "/mash_init" ^ serial
   349                  |> tap (Path.explode #> Isabelle_System.mkdir)
   350   in
   351     write_file write_access (in_dir ^ "/mash_accessibility");
   352     write_file write_feats (in_dir ^ "/mash_features");
   353     write_file write_deps (in_dir ^ "/mash_dependencies");
   354     run_mash ctxt overlord info false
   355              ("--init --inputDir " ^ in_dir ^
   356               and_rm_files overlord " -r" [in_dir])
   357   end
   358 
   359 fun run_mash_commands ctxt overlord save max_suggs write_cmds read_suggs =
   360   let
   361     val info as (temp_dir, serial) = mash_info overlord
   362     val sugg_file = temp_dir ^ "/mash_suggs" ^ serial
   363     val cmd_file = temp_dir ^ "/mash_commands" ^ serial
   364   in
   365     write_file ([], K "") sugg_file;
   366     write_file write_cmds cmd_file;
   367     run_mash ctxt overlord info false
   368              ("--inputFile " ^ cmd_file ^ " --predictions " ^ sugg_file ^
   369               " --numberOfPredictions " ^ string_of_int max_suggs ^
   370               (if save then " --saveModel" else "") ^
   371               (and_rm_files overlord "" [sugg_file, cmd_file]));
   372     read_suggs (fn () => File.read_lines (Path.explode sugg_file))
   373   end
   374 
   375 fun str_of_update (name, parents, feats, deps) =
   376   "! " ^ escape_meta name ^ ": " ^ escape_metas parents ^ "; " ^
   377   escape_metas feats ^ "; " ^ escape_metas deps ^ "\n"
   378 
   379 fun str_of_query (parents, feats) =
   380   "? " ^ escape_metas parents ^ "; " ^ escape_metas feats
   381 
   382 fun init_str_of_update get (upd as (name, _, _, _)) =
   383   escape_meta name ^ ": " ^ escape_metas (get upd) ^ "\n"
   384 
   385 fun mash_CLEAR ctxt =
   386   let val path = mash_state_dir () |> Path.explode in
   387     trace_msg ctxt (K "MaSh CLEAR");
   388     File.fold_dir (fn file => fn () =>
   389                       File.rm (Path.append path (Path.basic file)))
   390                   path ()
   391   end
   392 
   393 fun mash_INIT ctxt _ [] = mash_CLEAR ctxt
   394   | mash_INIT ctxt overlord upds =
   395     (trace_msg ctxt (fn () => "MaSh INIT " ^
   396          elide_string 1000 (space_implode " " (map #1 upds)));
   397      run_mash_init ctxt overlord (upds, init_str_of_update #2)
   398                    (upds, init_str_of_update #3) (upds, init_str_of_update #4))
   399 
   400 fun mash_ADD _ _ [] = ()
   401   | mash_ADD ctxt overlord upds =
   402     (trace_msg ctxt (fn () => "MaSh ADD " ^
   403          elide_string 1000 (space_implode " " (map #1 upds)));
   404      run_mash_commands ctxt overlord true 0 (upds, str_of_update) (K ()))
   405 
   406 fun mash_QUERY ctxt overlord max_suggs (query as (_, feats)) =
   407   (trace_msg ctxt (fn () => "MaSh QUERY " ^ space_implode " " feats);
   408    run_mash_commands ctxt overlord false  max_suggs
   409        ([query], str_of_query)
   410        (fn suggs => snd (extract_query (List.last (suggs ()))))
   411    handle List.Empty => [])
   412 
   413 
   414 (*** High-level communication with MaSh ***)
   415 
   416 fun try_graph ctxt when def f =
   417   f ()
   418   handle Graph.CYCLES (cycle :: _) =>
   419          (trace_msg ctxt (fn () =>
   420               "Cycle involving " ^ commas cycle ^ " when " ^ when); def)
   421        | Graph.UNDEF name =>
   422          (trace_msg ctxt (fn () =>
   423               "Unknown fact " ^ quote name ^ " when " ^ when); def)
   424 
   425 type mash_state =
   426   {thys : bool Symtab.table,
   427    fact_graph : unit Graph.T}
   428 
   429 val empty_state = {thys = Symtab.empty, fact_graph = Graph.empty}
   430 
   431 local
   432 
   433 fun mash_load _ (state as (true, _)) = state
   434   | mash_load ctxt _ =
   435     let val path = mash_state_path () in
   436       (true,
   437        case try File.read_lines path of
   438          SOME (comp_thys :: incomp_thys :: fact_lines) =>
   439          let
   440            fun add_thy comp thy = Symtab.update (thy, comp)
   441            fun add_edge_to name parent =
   442              Graph.default_node (parent, ())
   443              #> Graph.add_edge (parent, name)
   444            fun add_fact_line line =
   445              case extract_query line of
   446                ("", _) => I (* shouldn't happen *)
   447              | (name, parents) =>
   448                Graph.default_node (name, ())
   449                #> fold (add_edge_to name) parents
   450            val thys =
   451              Symtab.empty |> fold (add_thy true) (unescape_metas comp_thys)
   452                           |> fold (add_thy false) (unescape_metas incomp_thys)
   453            val fact_graph =
   454              try_graph ctxt "loading state" Graph.empty (fn () =>
   455                  Graph.empty |> fold add_fact_line fact_lines)
   456          in {thys = thys, fact_graph = fact_graph} end
   457        | _ => empty_state)
   458     end
   459 
   460 fun mash_save ({thys, fact_graph, ...} : mash_state) =
   461   let
   462     val path = mash_state_path ()
   463     val thys = Symtab.dest thys
   464     val line_for_thys = escape_metas o AList.find (op =) thys
   465     fun fact_line_for name parents =
   466       escape_meta name ^ ": " ^ escape_metas parents
   467     val append_fact = File.append path o suffix "\n" oo fact_line_for
   468   in
   469     File.write path (line_for_thys true ^ "\n" ^ line_for_thys false ^ "\n");
   470     Graph.fold (fn (name, ((), (parents, _))) => fn () =>
   471                    append_fact name (Graph.Keys.dest parents))
   472         fact_graph ()
   473   end
   474 
   475 val global_state =
   476   Synchronized.var "Sledgehammer_MaSh.global_state" (false, empty_state)
   477 
   478 in
   479 
   480 fun mash_map ctxt f =
   481   Synchronized.change global_state (mash_load ctxt ##> (f #> tap mash_save))
   482 
   483 fun mash_get ctxt =
   484   Synchronized.change_result global_state (mash_load ctxt #> `snd)
   485 
   486 fun mash_unlearn ctxt =
   487   Synchronized.change global_state (fn _ =>
   488       (mash_CLEAR ctxt; File.write (mash_state_path ()) "";
   489        (true, empty_state)))
   490 
   491 end
   492 
   493 fun mash_could_suggest_facts () = mash_home () <> ""
   494 fun mash_can_suggest_facts ctxt =
   495   not (Graph.is_empty (#fact_graph (mash_get ctxt)))
   496 
   497 fun parents_wrt_facts facts fact_graph =
   498   let
   499     val facts = [] |> fold (cons o nickname_of o snd) facts
   500     val tab = Symtab.empty |> fold (fn name => Symtab.update (name, ())) facts
   501     fun insert_not_seen seen name =
   502       not (member (op =) seen name) ? insert (op =) name
   503     fun parents_of _ parents [] = parents
   504       | parents_of seen parents (name :: names) =
   505         if Symtab.defined tab name then
   506           parents_of (name :: seen) (name :: parents) names
   507         else
   508           parents_of (name :: seen) parents
   509                      (Graph.Keys.fold (insert_not_seen seen)
   510                                       (Graph.imm_preds fact_graph name) names)
   511   in parents_of [] [] (Graph.maximals fact_graph) end
   512 
   513 (* Generate more suggestions than requested, because some might be thrown out
   514    later for various reasons and "meshing" gives better results with some
   515    slack. *)
   516 fun max_suggs_of max_facts = max_facts + Int.min (200, max_facts)
   517 
   518 fun is_fact_in_graph fact_graph (_, th) =
   519   can (Graph.get_node fact_graph) (nickname_of th)
   520 
   521 fun mash_suggest_facts ctxt ({overlord, ...} : params) prover max_facts hyp_ts
   522                        concl_t facts =
   523   let
   524     val thy = Proof_Context.theory_of ctxt
   525     val fact_graph = #fact_graph (mash_get ctxt)
   526     val parents = parents_wrt_facts facts fact_graph
   527     val feats = features_of ctxt prover thy General (concl_t :: hyp_ts)
   528     val suggs =
   529       if Graph.is_empty fact_graph then []
   530       else mash_QUERY ctxt overlord (max_suggs_of max_facts) (parents, feats)
   531     val selected = facts |> suggested_facts suggs
   532     val unknown = facts |> filter_out (is_fact_in_graph fact_graph)
   533   in (selected, unknown) end
   534 
   535 fun add_thys_for thy =
   536   let fun add comp thy = Symtab.update (Context.theory_name thy, comp) in
   537     add false thy #> fold (add true) (Theory.ancestors_of thy)
   538   end
   539 
   540 fun update_fact_graph ctxt (name, parents, feats, deps) (upds, graph) =
   541   let
   542     fun maybe_add_from from (accum as (parents, graph)) =
   543       try_graph ctxt "updating graph" accum (fn () =>
   544           (from :: parents, Graph.add_edge_acyclic (from, name) graph))
   545     val graph = graph |> Graph.default_node (name, ())
   546     val (parents, graph) = ([], graph) |> fold maybe_add_from parents
   547     val (deps, graph) = ([], graph) |> fold maybe_add_from deps
   548   in ((name, parents, feats, deps) :: upds, graph) end
   549 
   550 val pass1_learn_timeout_factor = 0.5
   551 
   552 fun mash_learn_proof ctxt ({provers, overlord, ...} : params) t facts used_ths =
   553   let
   554     val thy = Proof_Context.theory_of ctxt
   555     val prover = hd provers
   556     val name = ATP_Util.timestamp () ^ " " ^ serial_string () (* fresh enough *)
   557     val feats = features_of ctxt prover thy General [t]
   558     val deps = used_ths |> map nickname_of
   559   in
   560     mash_map ctxt (fn {thys, fact_graph} =>
   561         let
   562           val parents = parents_wrt_facts facts fact_graph
   563           val upds = [(name, parents, feats, deps)]
   564           val (upds, fact_graph) =
   565             ([], fact_graph) |> fold (update_fact_graph ctxt) upds
   566         in
   567           mash_ADD ctxt overlord upds;
   568           {thys = thys, fact_graph = fact_graph}
   569         end)
   570   end
   571 
   572 (* Too many dependencies is a sign that a decision procedure is at work. There
   573    isn't much too learn from such proofs. *)
   574 val max_dependencies = 10
   575 
   576 (* The timeout is understood in a very slack fashion. *)
   577 fun mash_learn_thy ctxt ({provers, verbose, overlord, ...} : params) thy timeout
   578                    facts =
   579   let
   580     val timer = Timer.startRealTimer ()
   581     val prover = hd provers
   582     fun timed_out frac =
   583       Time.> (Timer.checkRealTimer timer, time_mult frac timeout)
   584     val {fact_graph, ...} = mash_get ctxt
   585     val new_facts =
   586       facts |> filter_out (is_fact_in_graph fact_graph)
   587             |> sort (thm_ord o pairself snd)
   588   in
   589     if null new_facts then
   590       ""
   591     else
   592       let
   593         val ths = facts |> map snd
   594         val all_names =
   595           ths |> filter_out is_likely_tautology_or_too_meta
   596               |> map (rpair () o nickname_of)
   597               |> Symtab.make
   598         fun trim_deps deps = if length deps > max_dependencies then [] else deps
   599         fun do_fact _ (accum as (_, true)) = accum
   600           | do_fact ((_, (_, status)), th) ((parents, upds), false) =
   601             let
   602               val name = nickname_of th
   603               val feats =
   604                 features_of ctxt prover (theory_of_thm th) status [prop_of th]
   605               val deps = isabelle_dependencies_of all_names th |> trim_deps
   606               val upd = (name, parents, feats, deps)
   607             in (([name], upd :: upds), timed_out pass1_learn_timeout_factor) end
   608         val parents = parents_wrt_facts facts fact_graph
   609         val ((_, upds), _) =
   610           ((parents, []), false) |> fold do_fact new_facts |>> apsnd rev
   611         val n = length upds
   612         fun trans {thys, fact_graph} =
   613           let
   614             val mash_INIT_or_ADD =
   615               if Graph.is_empty fact_graph then mash_INIT else mash_ADD
   616             val (upds, fact_graph) =
   617               ([], fact_graph) |> fold (update_fact_graph ctxt) upds
   618           in
   619             (mash_INIT_or_ADD ctxt overlord (rev upds);
   620              {thys = thys |> add_thys_for thy, fact_graph = fact_graph})
   621           end
   622       in
   623         mash_map ctxt trans;
   624         if verbose then
   625           "Learned " ^ string_of_int n ^ " proof" ^ plural_s n ^
   626           (if verbose then
   627              " in " ^ string_from_time (Timer.checkRealTimer timer)
   628            else
   629              "") ^ "."
   630         else
   631           ""
   632       end
   633   end
   634 
   635 fun mash_learn ctxt params =
   636   let
   637     val thy = Proof_Context.theory_of ctxt
   638     val _ = Output.urgent_message "MaShing..."
   639     val css_table = Sledgehammer_Fact.clasimpset_rule_table_of ctxt
   640     val facts = all_facts_of thy css_table
   641   in
   642      mash_learn_thy ctxt params thy infinite_timeout facts
   643      |> (fn "" => "Learned nothing." | msg => msg)
   644      |> Output.urgent_message
   645   end
   646 
   647 (* The threshold should be large enough so that MaSh doesn't kick in for Auto
   648    Sledgehammer and Try. *)
   649 val min_secs_for_learning = 15
   650 val learn_timeout_factor = 2.0
   651 
   652 fun relevant_facts ctxt (params as {learn, fact_filter, timeout, ...}) prover
   653         max_facts ({add, only, ...} : fact_override) hyp_ts concl_t facts =
   654   if not (subset (op =) (the_list fact_filter, fact_filters)) then
   655     error ("Unknown fact filter: " ^ quote (the fact_filter) ^ ".")
   656   else if only then
   657     facts
   658   else if max_facts <= 0 orelse null facts then
   659     []
   660   else
   661     let
   662       val thy = Proof_Context.theory_of ctxt
   663       fun maybe_learn () =
   664         if not learn orelse Async_Manager.has_running_threads MaShN then
   665           ()
   666         else if Time.toSeconds timeout >= min_secs_for_learning then
   667           let
   668             val soft_timeout = time_mult learn_timeout_factor timeout
   669             val hard_timeout = time_mult 4.0 soft_timeout
   670             val birth_time = Time.now ()
   671             val death_time = Time.+ (birth_time, hard_timeout)
   672             val desc = ("machine learner for Sledgehammer", "")
   673           in
   674             Async_Manager.launch MaShN birth_time death_time desc
   675                 (fn () =>
   676                     (true, mash_learn_thy ctxt params thy soft_timeout facts))
   677           end
   678         else
   679           ()
   680       val fact_filter =
   681         case fact_filter of
   682           SOME ff => (() |> ff <> mepoN ? maybe_learn; ff)
   683         | NONE =>
   684           if mash_can_suggest_facts ctxt then (maybe_learn (); meshN)
   685           else if mash_could_suggest_facts () then (maybe_learn (); mepoN)
   686           else mepoN
   687       val add_ths = Attrib.eval_thms ctxt add
   688       fun prepend_facts ths accepts =
   689         ((facts |> filter (member Thm.eq_thm_prop ths o snd)) @
   690          (accepts |> filter_out (member Thm.eq_thm_prop ths o snd)))
   691         |> take max_facts
   692       fun iter () =
   693         iterative_relevant_facts ctxt params prover max_facts NONE hyp_ts
   694                                  concl_t facts
   695       fun mash () =
   696         mash_suggest_facts ctxt params prover max_facts hyp_ts concl_t facts
   697       val mess =
   698         [] |> (if fact_filter <> mashN then cons (iter (), []) else I)
   699            |> (if fact_filter <> mepoN then cons (mash ()) else I)
   700     in
   701       mesh_facts max_facts mess
   702       |> not (null add_ths) ? prepend_facts add_ths
   703     end
   704 
   705 fun kill_learners () = Async_Manager.kill_threads MaShN "learner"
   706 fun running_learners () = Async_Manager.running_threads MaShN "learner"
   707 
   708 end;