merged
authorwenzelm
Fri, 23 Aug 2013 21:59:29 +0200
changeset 53173 b881bee69d3a
parent 53159 a5805fe4e91c (diff)
parent 53172 31e24d6ff1ea (current diff)
child 53174 71a2702da5e0
merged
src/HOL/Spec_Check/Examples.thy
src/HOL/Spec_Check/README
src/HOL/Spec_Check/Spec_Check.thy
src/HOL/Spec_Check/base_generator.ML
src/HOL/Spec_Check/gen_construction.ML
src/HOL/Spec_Check/generator.ML
src/HOL/Spec_Check/output_style.ML
src/HOL/Spec_Check/property.ML
src/HOL/Spec_Check/random.ML
src/HOL/Spec_Check/spec_check.ML
src/Tools/WWW_Find/ROOT
--- a/src/HOL/TPTP/mash_export.ML	Fri Aug 23 20:53:00 2013 +0200
+++ b/src/HOL/TPTP/mash_export.ML	Fri Aug 23 21:59:29 2013 +0200
@@ -189,7 +189,7 @@
                   |> rev
                   |> weight_facts_steeply
                   |> map extra_features_of
-                  |> rpair goal_feats |-> fold (union (op = o pairself fst))
+                  |> rpair goal_feats |-> fold (union (eq_fst (op =)))
               in
                 "? " ^ string_of_int max_suggs ^ " # " ^
                 encode_str name ^ ": " ^ encode_strs parents ^ "; " ^
--- a/src/HOL/Tools/Sledgehammer/MaSh/src/mash.py	Fri Aug 23 20:53:00 2013 +0200
+++ b/src/HOL/Tools/Sledgehammer/MaSh/src/mash.py	Fri Aug 23 21:59:29 2013 +0200
@@ -103,30 +103,27 @@
         received = communicate(data,args.host,args.port)
         logger.info(received)     
     
-    if args.inputFile == None:
-        return
-    logger.debug('Using the following settings: %s',args)
-    # IO Streams
-    OS = open(args.predictions,'w')
-    IS = open(args.inputFile,'r')
-    lineCount = 0
-    for line in IS:
-        lineCount += 1
-        if lineCount % 100 == 0:
-            logger.info('On line %s', lineCount)
-        #if lineCount == 50: ###
-        #    break       
-        received = communicate(line,args.host,args.port)
-        if not received == '':
-            OS.write('%s\n' % received)
-    OS.close()
-    IS.close()
+    if not args.inputFile == None:
+        logger.debug('Using the following settings: %s',args)
+        # IO Streams
+        OS = open(args.predictions,'w')
+        IS = open(args.inputFile,'r')
+        lineCount = 0
+        for line in IS:
+            lineCount += 1
+            if lineCount % 100 == 0:
+                logger.info('On line %s', lineCount)
+            received = communicate(line,args.host,args.port)
+            if not received == '':
+                OS.write('%s\n' % received)
+        OS.close()
+        IS.close()
 
     # Statistics
     if args.statistics:
         received = communicate('avgStats',args.host,args.port)
         logger.info(received)
-    elif args.saveModels:
+    if args.saveModels:
         communicate('save',args.host,args.port)
 
 
--- a/src/HOL/Tools/Sledgehammer/MaSh/src/mashOld.py	Fri Aug 23 20:53:00 2013 +0200
+++ /dev/null	Thu Jan 01 00:00:00 1970 +0000
@@ -1,329 +0,0 @@
-#!/usr/bin/python
-#     Title:      HOL/Tools/Sledgehammer/MaSh/src/mash.py
-#     Author:     Daniel Kuehlwein, ICIS, Radboud University Nijmegen
-#     Copyright   2012
-#
-# Entry point for MaSh (Machine Learning for Sledgehammer).
-
-'''
-MaSh - Machine Learning for Sledgehammer
-
-MaSh allows to use different machine learning algorithms to predict relevant fact for Sledgehammer.
-
-Created on July 12, 2012
-
-@author: Daniel Kuehlwein
-'''
-
-import logging,datetime,string,os,sys
-from argparse import ArgumentParser,RawDescriptionHelpFormatter
-from time import time
-from stats import Statistics
-from theoryStats import TheoryStatistics
-from theoryModels import TheoryModels
-from dictionaries import Dictionaries
-#from fullNaiveBayes import NBClassifier
-from sparseNaiveBayes import sparseNBClassifier
-from snow import SNoW
-from predefined import Predefined
-
-# Set up command-line parser
-parser = ArgumentParser(description='MaSh - Machine Learning for Sledgehammer.  \n\n\
-MaSh allows to use different machine learning algorithms to predict relevant facts for Sledgehammer.\n\n\
---------------- Example Usage ---------------\n\
-First initialize:\n./mash.py -l test.log -o ../tmp/ --init --inputDir ../data/Jinja/ \n\
-Then create predictions:\n./mash.py -i ../data/Jinja/mash_commands -p ../data/Jinja/mash_suggestions -l test.log -o ../tmp/ --statistics\n\
-\n\n\
-Author: Daniel Kuehlwein, July 2012',formatter_class=RawDescriptionHelpFormatter)
-parser.add_argument('-i','--inputFile',help='File containing all problems to be solved.')
-parser.add_argument('-o','--outputDir', default='../tmp/',help='Directory where all created files are stored. Default=../tmp/.')
-parser.add_argument('-p','--predictions',default='../tmp/%s.predictions' % datetime.datetime.now(),
-                    help='File where the predictions stored. Default=../tmp/dateTime.predictions.')
-parser.add_argument('--numberOfPredictions',default=200,help="Number of premises to write in the output. Default=200.",type=int)
-
-parser.add_argument('--init',default=False,action='store_true',help="Initialize Mash. Requires --inputDir to be defined. Default=False.")
-parser.add_argument('--inputDir',default='../data/20121212/Jinja/',\
-                    help='Directory containing all the input data. MaSh expects the following files: mash_features,mash_dependencies,mash_accessibility')
-parser.add_argument('--depFile', default='mash_dependencies',
-                    help='Name of the file with the premise dependencies. The file must be in inputDir. Default = mash_dependencies')
-parser.add_argument('--saveModel',default=False,action='store_true',help="Stores the learned Model at the end of a prediction run. Default=False.")
-
-parser.add_argument('--learnTheories',default=False,action='store_true',help="Uses a two-lvl prediction mode. First the theories, then the premises. Default=False.")
-# Theory Parameters
-parser.add_argument('--theoryDefValPos',default=-7.5,help="Default value for positive unknown features. Default=-7.5.",type=float)
-parser.add_argument('--theoryDefValNeg',default=-10.0,help="Default value for negative unknown features. Default=-15.0.",type=float)
-parser.add_argument('--theoryPosWeight',default=2.0,help="Weight value for positive features. Default=2.0.",type=float)
-
-parser.add_argument('--nb',default=False,action='store_true',help="Use Naive Bayes for learning. This is the default learning method.")
-# NB Parameters
-parser.add_argument('--NBDefaultPriorWeight',default=20.0,help="Initializes classifiers with value * p |- p. Default=20.0.",type=float)
-parser.add_argument('--NBDefVal',default=-15.0,help="Default value for unknown features. Default=-15.0.",type=float)
-parser.add_argument('--NBPosWeight',default=10.0,help="Weight value for positive features. Default=10.0.",type=float)
-# TODO: Rename to sineFeatures
-parser.add_argument('--sineFeatures',default=False,action='store_true',help="Uses a SInE like prior for premise lvl predictions. Default=False.")
-parser.add_argument('--sineWeight',default=0.5,help="How much the SInE prior is weighted. Default=0.5.",type=float)
-
-parser.add_argument('--snow',default=False,action='store_true',help="Use SNoW's naive bayes instead of Naive Bayes for learning.")
-parser.add_argument('--predef',help="Use predefined predictions. Used only for comparison with the actual learning. Argument is the filename of the predictions.")
-parser.add_argument('--statistics',default=False,action='store_true',help="Create and show statistics for the top CUTOFF predictions.\
-                    WARNING: This will make the program a lot slower! Default=False.")
-parser.add_argument('--saveStats',default=None,help="If defined, stores the statistics in the filename provided.")
-parser.add_argument('--cutOff',default=500,help="Option for statistics. Only consider the first cutOff predictions. Default=500.",type=int)
-parser.add_argument('-l','--log', default='../tmp/%s.log' % datetime.datetime.now(), help='Log file name. Default=../tmp/dateTime.log')
-parser.add_argument('-q','--quiet',default=False,action='store_true',help="If enabled, only print warnings. Default=False.")
-parser.add_argument('--modelFile', default='../tmp/model.pickle', help='Model file name. Default=../tmp/model.pickle')
-parser.add_argument('--dictsFile', default='../tmp/dict.pickle', help='Dict file name. Default=../tmp/dict.pickle')
-parser.add_argument('--theoryFile', default='../tmp/theory.pickle', help='Model file name. Default=../tmp/theory.pickle')
-
-def mash(argv = sys.argv[1:]):
-    # Initializing command-line arguments
-    args = parser.parse_args(argv)
-    
-    # Set up logging
-    logging.basicConfig(level=logging.DEBUG,
-                        format='%(asctime)s %(name)-12s %(levelname)-8s %(message)s',
-                        datefmt='%d-%m %H:%M:%S',
-                        filename=args.log,
-                        filemode='w')    
-    logger = logging.getLogger('main.py')
-    
-    """
-    # remove old handler for tester
-    # TODO: Comment out for Jasmins version. This crashes python 2.6.1
-    logger.root.handlers[0].stream.close()
-    logger.root.removeHandler(logger.root.handlers[0])
-    file_handler = logging.FileHandler(args.log)
-    file_handler.setLevel(logging.DEBUG)
-    formatter = logging.Formatter('%(asctime)s %(name)-12s %(levelname)-8s %(message)s')
-    file_handler.setFormatter(formatter)
-    logger.root.addHandler(file_handler)
-    #"""
-    if args.quiet:
-        logger.setLevel(logging.WARNING)
-        #console.setLevel(logging.WARNING)
-    else:
-        console = logging.StreamHandler(sys.stdout)
-        console.setLevel(logging.INFO)
-        formatter = logging.Formatter('# %(message)s')
-        console.setFormatter(formatter)
-        logging.getLogger('').addHandler(console)
-        
-    if not os.path.exists(args.outputDir):
-        os.makedirs(args.outputDir)
-
-    logger.info('Using the following settings: %s',args)
-    # Pick algorithm
-    if args.nb:
-        logger.info('Using sparse Naive Bayes for learning.')
-        model = sparseNBClassifier(args.NBDefaultPriorWeight,args.NBPosWeight,args.NBDefVal)
-    elif args.snow:
-        logger.info('Using naive bayes (SNoW) for learning.')
-        model = SNoW()
-    elif args.predef:
-        logger.info('Using predefined predictions.')
-        model = Predefined(args.predef)
-    else:
-        logger.info('No algorithm specified. Using sparse Naive Bayes.')
-        model = sparseNBClassifier(args.NBDefaultPriorWeight,args.NBPosWeight,args.NBDefVal)
-
-    # Initializing model
-    if args.init:
-        logger.info('Initializing Model.')
-        startTime = time()
-
-        # Load all data
-        dicts = Dictionaries()
-        dicts.init_all(args)
-        
-        # Create Model
-        trainData = dicts.featureDict.keys()
-        model.initializeModel(trainData,dicts)
-
-        if args.learnTheories:
-            depFile = os.path.join(args.inputDir,args.depFile)
-            theoryModels = TheoryModels(args.theoryDefValPos,args.theoryDefValNeg,args.theoryPosWeight)
-            theoryModels.init(depFile,dicts)
-            theoryModels.save(args.theoryFile)
-            
-        model.save(args.modelFile)
-        dicts.save(args.dictsFile)
-
-        logger.info('All Done. %s seconds needed.',round(time()-startTime,2))
-        return 0
-    # Create predictions and/or update model
-    else:
-        lineCounter = 1
-        statementCounter = 1
-        computeStats = False
-        dicts = Dictionaries()
-        theoryModels = TheoryModels(args.theoryDefValPos,args.theoryDefValNeg,args.theoryPosWeight)
-        # Load Files
-        if os.path.isfile(args.dictsFile):
-            #logger.info('Loading Dictionaries')
-            #startTime = time()
-            dicts.load(args.dictsFile)            
-            #logger.info('Done %s',time()-startTime)
-        if os.path.isfile(args.modelFile):
-            #logger.info('Loading Model')
-            #startTime = time()
-            model.load(args.modelFile)            
-            #logger.info('Done %s',time()-startTime)
-        if os.path.isfile(args.theoryFile) and args.learnTheories:
-            #logger.info('Loading Theory Models')
-            #startTime = time()
-            theoryModels.load(args.theoryFile)
-            #logger.info('Done %s',time()-startTime)
-        logger.info('All loading completed')
-
-        # IO Streams
-        OS = open(args.predictions,'w')
-        IS = open(args.inputFile,'r')
-
-        # Statistics
-        if args.statistics:
-            stats = Statistics(args.cutOff)
-            if args.learnTheories:
-                theoryStats = TheoryStatistics()
-
-        predictions = None
-        predictedTheories = None
-        #Reading Input File
-        for line in IS:
-#           try:
-            if True:
-                if line.startswith('!'):
-                    problemId = dicts.parse_fact(line)    
-                    # Statistics
-                    if args.statistics and computeStats:
-                        computeStats = False
-                        # Assume '!' comes after '?'
-                        if args.predef:
-                            predictions = model.predict(problemId)
-                        if args.learnTheories:
-                            tmp = [dicts.idNameDict[x] for x in dicts.dependenciesDict[problemId]]
-                            usedTheories = set([x.split('.')[0] for x in tmp]) 
-                            theoryStats.update((dicts.idNameDict[problemId]).split('.')[0],predictedTheories,usedTheories,len(theoryModels.accessibleTheories))                        
-                        stats.update(predictions,dicts.dependenciesDict[problemId],statementCounter)
-                        if not stats.badPreds == []:
-                            bp = string.join([str(dicts.idNameDict[x]) for x in stats.badPreds], ',')
-                            logger.debug('Bad predictions: %s',bp)
-
-                    statementCounter += 1
-                    # Update Dependencies, p proves p
-                    dicts.dependenciesDict[problemId] = [problemId]+dicts.dependenciesDict[problemId]
-                    if args.learnTheories:
-                        theoryModels.update(problemId,dicts.featureDict[problemId],dicts.dependenciesDict[problemId],dicts)
-                    if args.snow:
-                        model.update(problemId,dicts.featureDict[problemId],dicts.dependenciesDict[problemId],dicts)
-                    else:
-                        model.update(problemId,dicts.featureDict[problemId],dicts.dependenciesDict[problemId])
-                elif line.startswith('p'):
-                    # Overwrite old proof.
-                    problemId,newDependencies = dicts.parse_overwrite(line)
-                    newDependencies = [problemId]+newDependencies
-                    model.overwrite(problemId,newDependencies,dicts)
-                    if args.learnTheories:
-                        theoryModels.overwrite(problemId,newDependencies,dicts)
-                    dicts.dependenciesDict[problemId] = newDependencies
-                elif line.startswith('?'):               
-                    startTime = time()
-                    computeStats = True
-                    if args.predef:
-                        continue
-                    name,features,accessibles,hints = dicts.parse_problem(line)  
-                        
-                    # Create predictions
-                    logger.info('Starting computation for problem on line %s',lineCounter)
-                    # Update Models with hints
-                    if not hints == []:
-                        if args.learnTheories:
-                            accessibleTheories = set([(dicts.idNameDict[x]).split('.')[0] for x in accessibles])
-                            theoryModels.update_with_acc('hints',features,hints,dicts,accessibleTheories)
-                        if args.snow:
-                            pass
-                        else:
-                            model.update('hints',features,hints)
-
-                    # Predict premises
-                    if args.learnTheories:
-                        predictedTheories,accessibles = theoryModels.predict(features,accessibles,dicts)
-
-                    # Add additional features on premise lvl if sine is enabled
-                    if args.sineFeatures:
-                        origFeatures = [f for f,_w in features]
-                        secondaryFeatures = []
-                        for f in origFeatures:
-                            if dicts.featureCountDict[f] == 1:
-                                continue
-                            triggeredFormulas = dicts.featureTriggeredFormulasDict[f]                                
-                            for formula in triggeredFormulas: 
-                                tFeatures = dicts.triggerFeaturesDict[formula]                                
-                                #tFeatures = [ff for ff,_fw in dicts.featureDict[formula]]
-                                newFeatures = set(tFeatures).difference(secondaryFeatures+origFeatures)
-                            for fNew in newFeatures:
-                                secondaryFeatures.append((fNew,args.sineWeight))
-                        predictionsFeatures = features+secondaryFeatures
-                    else:
-                        predictionsFeatures = features                    
-                    predictions,predictionValues = model.predict(predictionsFeatures,accessibles,dicts)
-                    assert len(predictions) == len(predictionValues)
-                    
-                    # Delete hints
-                    if not hints == []:
-                        if args.learnTheories:
-                            theoryModels.delete('hints',features,hints,dicts)
-                        if args.snow:
-                            pass
-                        else:
-                            model.delete('hints',features,hints)
-
-                    logger.info('Done. %s seconds needed.',round(time()-startTime,2))
-                    # Output        
-                    predictionNames = [str(dicts.idNameDict[p]) for p in predictions[:args.numberOfPredictions]]
-                    predictionValues = [str(x) for x in predictionValues[:args.numberOfPredictions]]
-                    predictionsStringList = ['%s=%s' % (predictionNames[i],predictionValues[i]) for i in range(len(predictionNames))]
-                    predictionsString = string.join(predictionsStringList,' ')
-                    outString = '%s: %s' % (name,predictionsString)
-                    OS.write('%s\n' % outString)
-                else:
-                    logger.warning('Unspecified input format: \n%s',line)
-                    sys.exit(-1)
-                lineCounter += 1
-            """
-            except:
-                logger.warning('An error occurred on line %s .',line)
-                lineCounter += 1
-                continue
-            """
-        OS.close()
-        IS.close()
-
-        # Statistics
-        if args.statistics:
-            if args.learnTheories:
-                theoryStats.printAvg()
-            stats.printAvg()
-
-        # Save
-        if args.saveModel:
-            model.save(args.modelFile)
-            if args.learnTheories:
-                theoryModels.save(args.theoryFile)
-        dicts.save(args.dictsFile)
-        if not args.saveStats == None:
-            if args.learnTheories:
-                theoryStatsFile = os.path.join(args.outputDir,'theoryStats')
-                theoryStats.save(theoryStatsFile)
-            statsFile = os.path.join(args.outputDir,args.saveStats)
-            stats.save(statsFile)
-    return 0
-
-if __name__ == '__main__':
-    # Cezary Auth  
-    args = ['--statistics', '--init', '--inputDir', '../data/20130118/Jinja', '--log', '../tmp/auth.log', '--theoryFile', '../tmp/t0', '--modelFile', '../tmp/m0', '--dictsFile', '../tmp/d0','--NBDefaultPriorWeight', '20.0', '--NBDefVal', '-15.0', '--NBPosWeight', '10.0']
-    mash(args)
-    args = ['-i', '../data/20130118/Jinja/mash_commands', '-p', '../tmp/auth.pred0', '--statistics', '--cutOff', '500', '--log', '../tmp/auth.log','--modelFile', '../tmp/m0', '--dictsFile', '../tmp/d0']
-    mash(args)   
-
-    #sys.exit(mash(args))
-    sys.exit(mash())
--- a/src/HOL/Tools/Sledgehammer/MaSh/src/stats.py	Fri Aug 23 20:53:00 2013 +0200
+++ b/src/HOL/Tools/Sledgehammer/MaSh/src/stats.py	Fri Aug 23 21:59:29 2013 +0200
@@ -123,6 +123,8 @@
         # HACK FOR PAPER
         assert len(self.aucData) == len(self.recall100Median)
         nrDataPoints = len(self.aucData)
+        if nrDataPoints == 0:
+            return "No data points"
         if nrDataPoints % 2 == 1:
             medianAUC = sorted(self.aucData)[nrDataPoints/2 + 1]
         else:
--- a/src/HOL/Tools/Sledgehammer/sledgehammer_mash.ML	Fri Aug 23 20:53:00 2013 +0200
+++ b/src/HOL/Tools/Sledgehammer/sledgehammer_mash.ML	Fri Aug 23 21:59:29 2013 +0200
@@ -151,7 +151,7 @@
    xs |> chunk_list 500 |> List.app (File.append path o implode o map f))
   handle IO.Io _ => ()
 
-fun run_mash_tool ctxt overlord extra_args write_cmds read_suggs =
+fun run_mash_tool ctxt overlord extra_args background write_cmds read_suggs =
   let
     val (temp_dir, serial) =
       if overlord then (getenv "ISABELLE_HOME_USER", "")
@@ -172,7 +172,8 @@
       " --dictsFile=" ^ model_dir ^ "/dict.pickle" ^
       " --log " ^ log_file ^ " " ^ core ^
       (if extra_args = [] then "" else " " ^ space_implode " " extra_args) ^
-      " >& " ^ err_file
+      " >& " ^ err_file ^
+      (if background then " &" else "")
     fun run_on () =
       (Isabelle_System.bash command
        |> tap (fn _ => trace_msg ctxt (fn () =>
@@ -254,7 +255,12 @@
 struct
 
 fun shutdown ctxt overlord =
-  run_mash_tool ctxt overlord [shutdown_server_arg] ([], K "") (K ())
+  (trace_msg ctxt (K "MaSh shutdown");
+   run_mash_tool ctxt overlord [shutdown_server_arg] true ([], K "") (K ()))
+
+fun save ctxt overlord =
+  (trace_msg ctxt (K "MaSh save");
+   run_mash_tool ctxt overlord [save_models_arg] true ([], K "") (K ()))
 
 fun unlearn ctxt overlord =
   let val path = mash_model_dir () in
@@ -270,19 +276,19 @@
   | learn ctxt overlord learns =
     (trace_msg ctxt (fn () => "MaSh learn " ^
          elide_string 1000 (space_implode " " (map #1 learns)));
-     run_mash_tool ctxt overlord [save_models_arg] (learns, str_of_learn)
+     run_mash_tool ctxt overlord [] false (learns, str_of_learn)
                    (K ()))
 
 fun relearn _ _ [] = ()
   | relearn ctxt overlord relearns =
     (trace_msg ctxt (fn () => "MaSh relearn " ^
          elide_string 1000 (space_implode " " (map #1 relearns)));
-     run_mash_tool ctxt overlord [save_models_arg] (relearns, str_of_relearn)
+     run_mash_tool ctxt overlord [] false (relearns, str_of_relearn)
                    (K ()))
 
 fun query ctxt overlord max_suggs (query as (_, _, _, feats)) =
   (trace_msg ctxt (fn () => "MaSh query " ^ encode_features feats);
-   run_mash_tool ctxt overlord [] ([query], str_of_query max_suggs)
+   run_mash_tool ctxt overlord [] false ([query], str_of_query max_suggs)
        (fn suggs =>
            case suggs () of
              [] => []
@@ -359,8 +365,8 @@
      | _ => NONE)
   | _ => NONE
 
-fun load _ _ (state as (true, _)) = state
-  | load ctxt overlord _ =
+fun load_state _ _ (state as (true, _)) = state
+  | load_state ctxt overlord _ =
     let val path = mash_state_file () in
       (true,
        case try File.read_lines path of
@@ -394,8 +400,8 @@
        | _ => empty_state)
     end
 
-fun save _ (state as {dirty = SOME [], ...}) = state
-  | save ctxt {access_G, num_known_facts, dirty} =
+fun save_state _ (state as {dirty = SOME [], ...}) = state
+  | save_state ctxt {access_G, num_known_facts, dirty} =
     let
       fun str_of_entry (name, parents, kind) =
         str_of_proof_kind kind ^ " " ^ encode_str name ^ ": " ^
@@ -424,12 +430,13 @@
 in
 
 fun map_state ctxt overlord f =
-  Synchronized.change global_state (load ctxt overlord ##> (f #> save ctxt))
+  Synchronized.change global_state
+                      (load_state ctxt overlord ##> (f #> save_state ctxt))
   handle FILE_VERSION_TOO_NEW () => ()
 
 fun peek_state ctxt overlord f =
   Synchronized.change_result global_state
-      (perhaps (try (load ctxt overlord)) #> `snd #>> f)
+      (perhaps (try (load_state ctxt overlord)) #> `snd #>> f)
 
 fun clear_state ctxt overlord =
   Synchronized.change global_state (fn _ =>
@@ -513,14 +520,11 @@
             |> map snd |> take max_facts
     end
 
+fun free_feature_of s = ("f" ^ s, 40.0 (* FUDGE *))
 fun thy_feature_of s = ("y" ^ s, 8.0 (* FUDGE *))
-fun free_feature_of s = ("f" ^ s, 40.0 (* FUDGE *))
 fun type_feature_of s = ("t" ^ s, 4.0 (* FUDGE *))
 fun class_feature_of s = ("s" ^ s, 1.0 (* FUDGE *))
-fun status_feature_of status = (string_of_status status, 2.0 (* FUDGE *))
-val local_feature = ("local", 8.0 (* FUDGE *))
-val lams_feature = ("lams", 2.0 (* FUDGE *))
-val skos_feature = ("skos", 2.0 (* FUDGE *))
+val local_feature = ("local", 16.0 (* FUDGE *))
 
 fun crude_theory_ord p =
   if Theory.subthy p then
@@ -609,7 +613,7 @@
               #> swap #> op ::
               #> subtract (op =) @{sort type} #> map massage_long_name
               #> map class_feature_of
-              #> union (op = o pairself fst)) S
+              #> union (eq_fst (op =))) S
 
     fun pattify_type 0 _ = []
       | pattify_type _ (Type (s, [])) =
@@ -625,13 +629,13 @@
       | pattify_type _ (TVar (_, S)) =
         maybe_singleton_str pat_tvar_prefix (crude_str_of_sort S)
     fun add_type_pat depth T =
-      union (op = o pairself fst) (map type_feature_of (pattify_type depth T))
+      union (eq_fst (op =)) (map type_feature_of (pattify_type depth T))
     fun add_type_pats 0 _ = I
       | add_type_pats depth t =
         add_type_pat depth t #> add_type_pats (depth - 1) t
     fun add_type T =
       add_type_pats type_max_depth T
-      #> fold_atyps_sorts (fn (_, S) => add_classes S) T
+      #> fold_atyps_sorts (add_classes o snd) T
     fun add_subtypes (T as Type (_, Ts)) = add_type T #> fold add_subtypes Ts
       | add_subtypes T = add_type T
 
@@ -672,8 +676,7 @@
                | (q, qw) => (p ^ "(" ^ q ^ ")", pw + qw)) ps qs
         end
       | pattify_term _ _ _ _ = []
-    fun add_term_pat Ts depth =
-      union (op = o pairself fst) o pattify_term Ts [] depth
+    fun add_term_pat Ts = union (eq_fst (op =)) oo pattify_term Ts []
     fun add_term_pats _ 0 _ = I
       | add_term_pats Ts depth t =
         add_term_pat Ts depth t #> add_term_pats Ts (depth - 1) t
@@ -695,21 +698,16 @@
         #> fold (add_subterms Ts) args
   in [] |> fold (add_subterms []) ts end
 
-fun is_exists (s, _) = (s = @{const_name Ex} orelse s = @{const_name Ex1})
-
 val term_max_depth = 2
-val type_max_depth = 2
+val type_max_depth = 1
 
 (* TODO: Generate type classes for types? *)
-fun features_of ctxt prover thy num_facts const_tab (scope, status) ts =
+fun features_of ctxt prover thy num_facts const_tab (scope, _) ts =
   let val thy_name = Context.theory_name thy in
     thy_feature_of thy_name ::
     term_features_of ctxt prover thy_name num_facts const_tab term_max_depth
                      type_max_depth ts
-    |> status <> General ? cons (status_feature_of status)
     |> scope <> Global ? cons local_feature
-    |> exists (not o is_lambda_free) ts ? cons lams_feature
-    |> exists (exists_Const is_exists) ts ? cons skos_feature
   end
 
 (* Too many dependencies is a sign that a decision procedure is at work. There
@@ -906,7 +904,7 @@
 
 val chained_feature_factor = 0.5
 val extra_feature_factor = 0.1
-val num_extra_feature_facts = 0 (* FUDGE *)
+val num_extra_feature_facts = 10 (* FUDGE *)
 
 (* FUDGE *)
 fun weight_of_proximity_fact rank =
@@ -927,7 +925,7 @@
 fun find_mash_suggestions _ _ [] _ _ raw_unknown = ([], raw_unknown)
   | find_mash_suggestions ctxt max_facts suggs facts chained raw_unknown =
     let
-      val inter_fact = inter (Thm.eq_thm_prop o pairself snd)
+      val inter_fact = inter (eq_snd Thm.eq_thm_prop)
       val raw_mash = find_suggested_facts ctxt facts suggs
       val proximate = take max_proximity_facts facts
       val unknown_chained = inter_fact raw_unknown chained
@@ -938,9 +936,9 @@
          (0.1 (* FUDGE *), (weight_facts_steeply raw_mash, raw_unknown))]
       val unknown =
         raw_unknown
-        |> fold (subtract (Thm.eq_thm_prop o pairself snd))
+        |> fold (subtract (eq_snd Thm.eq_thm_prop))
                 [unknown_chained, unknown_proximate]
-    in (mesh_facts (Thm.eq_thm_prop o pairself snd) max_facts mess, unknown) end
+    in (mesh_facts (eq_snd Thm.eq_thm_prop) max_facts mess, unknown) end
 
 fun add_const_counts t =
   fold (fn s => Symtab.map_default (s, 0) (Integer.add 1))
@@ -972,15 +970,15 @@
                 chained
                 |> map (rpair 1.0)
                 |> map (chained_or_extra_features_of chained_feature_factor)
-                |> rpair [] |-> fold (union (op = o pairself fst))
+                |> rpair [] |-> fold (union (eq_fst (op =)))
               val extra_feats =
                 facts
                 |> take (Int.max (0, num_extra_feature_facts - length chained))
                 |> weight_facts_steeply
                 |> map (chained_or_extra_features_of extra_feature_factor)
-                |> rpair [] |-> fold (union (op = o pairself fst))
+                |> rpair [] |-> fold (union (eq_fst (op =)))
               val feats =
-                fold (union (op = o pairself fst)) [chained_feats, extra_feats]
+                fold (union (eq_fst (op =))) [chained_feats, extra_feats]
                      goal_feats
               val hints =
                 chained |> filter (is_fact_in_graph access_G o snd)
@@ -1044,7 +1042,8 @@
                 used_ths |> filter (is_fact_in_graph access_G)
                          |> map nickname_of_thm
             in
-              MaSh.learn ctxt overlord [(name, parents, feats, deps)]
+              MaSh.learn ctxt overlord [(name, parents, feats, deps)];
+              MaSh.save ctxt overlord
             end);
         (true, "")
       end)
@@ -1056,7 +1055,7 @@
 
 (* The timeout is understood in a very relaxed fashion. *)
 fun mash_learn_facts ctxt (params as {debug, verbose, overlord, ...}) prover
-                     auto_level run_prover learn_timeout facts =
+                     save auto_level run_prover learn_timeout facts =
   let
     val timer = Timer.startRealTimer ()
     fun next_commit_time () =
@@ -1107,6 +1106,7 @@
             in
               MaSh.learn ctxt overlord (rev learns);
               MaSh.relearn ctxt overlord relearns;
+              if save then MaSh.save ctxt overlord else ();
               {access_G = access_G, num_known_facts = num_known_facts,
                dirty = dirty}
             end
@@ -1228,7 +1228,7 @@
     val num_facts = length facts
     val prover = hd provers
     fun learn auto_level run_prover =
-      mash_learn_facts ctxt params prover auto_level run_prover NONE facts
+      mash_learn_facts ctxt params prover true auto_level run_prover NONE facts
       |> Output.urgent_message
   in
     if run_prover then
@@ -1261,11 +1261,14 @@
 val mepo_weight = 0.5
 val mash_weight = 0.5
 
+val max_facts_to_learn_before_query = 100
+
 (* The threshold should be large enough so that MaSh doesn't kick in for Auto
    Sledgehammer and Try. *)
 val min_secs_for_learning = 15
 
-fun relevant_facts ctxt (params as {overlord, learn, fact_filter, timeout, ...})
+fun relevant_facts ctxt
+        (params as {overlord, blocking, learn, fact_filter, timeout, ...})
         prover max_facts ({add, only, ...} : fact_override) hyp_ts concl_t
         facts =
   if not (subset (op =) (the_list fact_filter, fact_filters)) then
@@ -1278,28 +1281,46 @@
     [("", [])]
   else
     let
-      fun maybe_learn () =
-        if learn andalso not (Async_Manager.has_running_threads MaShN) andalso
+      fun maybe_launch_thread () =
+        if not blocking andalso
+           not (Async_Manager.has_running_threads MaShN) andalso
            (timeout = NONE orelse
             Time.toSeconds (the timeout) >= min_secs_for_learning) then
           let
             val timeout = Option.map (time_mult learn_timeout_slack) timeout
           in
             launch_thread (timeout |> the_default one_day)
-                (fn () => (true, mash_learn_facts ctxt params prover 2 false
-                                                  timeout facts))
+                (fn () => (true, mash_learn_facts ctxt params prover true 2
+                                                  false timeout facts))
           end
         else
           ()
-      val effective_fact_filter =
+      fun maybe_learn () =
+        if learn then
+          let
+            val {access_G, num_known_facts, ...} = peek_state ctxt overlord I
+            val is_in_access_G = is_fact_in_graph access_G o snd
+          in
+            if length facts - num_known_facts <= max_facts_to_learn_before_query
+               andalso length (filter_out is_in_access_G facts)
+                       <= max_facts_to_learn_before_query then
+              (mash_learn_facts ctxt params prover false 2 false timeout facts
+               |> (fn "" => () | s => Output.urgent_message (MaShN ^ ": " ^ s));
+               true)
+            else
+              (maybe_launch_thread (); false)
+          end
+        else
+          false
+      val (save, effective_fact_filter) =
         case fact_filter of
-          SOME ff => (() |> ff <> mepoN ? maybe_learn; ff)
+          SOME ff => (ff <> mepoN andalso maybe_learn (), ff)
         | NONE =>
           if is_mash_enabled () then
-            (maybe_learn ();
+            (maybe_learn (),
              if mash_can_suggest_facts ctxt overlord then meshN else mepoN)
           else
-            mepoN
+            (false, mepoN)
       val add_ths = Attrib.eval_thms ctxt add
       fun in_add (_, th) = member Thm.eq_thm_prop add_ths th
       fun add_and_take accepts =
@@ -1327,9 +1348,10 @@
                else
                  I)
       val mesh =
-        mesh_facts (Thm.eq_thm_prop o pairself snd) max_facts mess
+        mesh_facts (eq_snd Thm.eq_thm_prop) max_facts mess
         |> add_and_take
     in
+      if save then MaSh.save ctxt overlord else ();
       case (fact_filter, mess) of
         (NONE, [(_, (mepo, _)), (_, (mash, _))]) =>
         [(meshN, mesh), (mepoN, mepo |> map fst |> add_and_take),
--- a/src/HOL/Tools/Sledgehammer/sledgehammer_mepo.ML	Fri Aug 23 20:53:00 2013 +0200
+++ b/src/HOL/Tools/Sledgehammer/sledgehammer_mepo.ML	Fri Aug 23 21:59:29 2013 +0200
@@ -311,10 +311,10 @@
   s = pseudo_abs_name orelse String.isPrefix pseudo_skolem_prefix s orelse
   String.isSuffix theory_const_suffix s
 
-fun fact_weight fudge stature const_tab relevant_consts chained_consts
-                fact_consts =
-  case fact_consts |> List.partition (pconst_hyper_mem I relevant_consts)
-                   ||> filter_out (pconst_hyper_mem swap relevant_consts) of
+fun fact_weight fudge stature const_tab rel_const_tab rel_const_iter_tab
+                chained_const_tab fact_consts =
+  case fact_consts |> List.partition (pconst_hyper_mem I rel_const_tab)
+                   ||> filter_out (pconst_hyper_mem swap rel_const_tab) of
     ([], _) => 0.0
   | (rel, irrel) =>
     if forall (forall (is_odd_const_name o fst)) [rel, irrel] then
@@ -327,7 +327,8 @@
         val irrel_weight =
           ~ (stature_bonus fudge stature)
           |> fold (curry (op +)
-                   o irrel_pconst_weight fudge const_tab chained_consts) irrel
+                   o irrel_pconst_weight fudge const_tab chained_const_tab)
+                  irrel
         val res = rel_weight / (rel_weight + irrel_weight)
       in if Real.isFinite res then res else 0.0 end
 
@@ -400,30 +401,36 @@
                             | _ => NONE)
     val chained_const_tab = Symtab.empty |> fold (add_pconsts true) chained_ts
     val goal_const_tab =
-      Symtab.empty |> fold (add_pconsts true) hyp_ts
-                   |> add_pconsts false concl_t
+      Symtab.empty
+      |> fold (add_pconsts true) hyp_ts
+      |> add_pconsts false concl_t
       |> (fn tab => if Symtab.is_empty tab then chained_const_tab else tab)
       |> fold (if_empty_replace_with_scope thy is_built_in_const facts)
               [Chained, Assum, Local]
-    fun iter j remaining_max thres rel_const_tab hopeless hopeful =
+    val goal_const_iter_tab = goal_const_tab |> Symtab.map (K (K ~1))
+    fun iter j remaining_max thres rel_const_tab rel_const_iter_tab hopeless
+             hopeful =
       let
         fun relevant [] _ [] =
             (* Nothing has been added this iteration. *)
             if j = 0 andalso thres >= ridiculous_threshold then
               (* First iteration? Try again. *)
               iter 0 max_facts (thres / threshold_divisor) rel_const_tab
-                   hopeless hopeful
+                   rel_const_iter_tab hopeless hopeful
             else
               []
           | relevant candidates rejects [] =
             let
               val (accepts, more_rejects) =
                 take_most_relevant ctxt max_facts remaining_max fudge candidates
+              val sps = maps (snd o fst) accepts;
               val rel_const_tab' =
-                rel_const_tab
-                |> fold (add_pconst_to_table false) (maps (snd o fst) accepts)
-              fun is_dirty (c, _) =
-                Symtab.lookup rel_const_tab' c <> Symtab.lookup rel_const_tab c
+                rel_const_tab |> fold (add_pconst_to_table false) sps
+              val rel_const_iter_tab' =
+                rel_const_iter_tab
+                |> fold (fn (s, _) => Symtab.default (s, j)) sps
+              fun is_dirty (s, _) =
+                Symtab.lookup rel_const_tab' s <> Symtab.lookup rel_const_tab s
               val (hopeful_rejects, hopeless_rejects) =
                  (rejects @ hopeless, ([], []))
                  |-> fold (fn (ax as (_, consts), old_weight) =>
@@ -441,7 +448,8 @@
               val remaining_max = remaining_max - length accepts
             in
               trace_msg ctxt (fn () => "New or updated constants: " ^
-                  commas (rel_const_tab' |> Symtab.dest
+                  commas (rel_const_tab'
+                          |> Symtab.dest
                           |> subtract (op =) (rel_const_tab |> Symtab.dest)
                           |> map string_of_hyper_pconst));
               map (fst o fst) accepts @
@@ -449,7 +457,7 @@
                  []
                else
                  iter (j + 1) remaining_max thres rel_const_tab'
-                      hopeless_rejects hopeful_rejects)
+                      rel_const_iter_tab' hopeless_rejects hopeful_rejects)
             end
           | relevant candidates rejects
                      (((ax as (((_, stature), _), fact_consts)), cached_weight)
@@ -458,8 +466,9 @@
               val weight =
                 case cached_weight of
                   SOME w => w
-                | NONE => fact_weight fudge stature const_tab rel_const_tab
-                                      chained_const_tab fact_consts
+                | NONE =>
+                  fact_weight fudge stature const_tab rel_const_tab
+                              rel_const_iter_tab chained_const_tab fact_consts
             in
               if weight >= thres then
                 relevant ((ax, weight) :: candidates) rejects hopeful
@@ -470,7 +479,8 @@
           trace_msg ctxt (fn () =>
               "ITERATION " ^ string_of_int j ^ ": current threshold: " ^
               Real.toString thres ^ ", constants: " ^
-              commas (rel_const_tab |> Symtab.dest
+              commas (rel_const_tab
+                      |> Symtab.dest
                       |> filter (curry (op <>) [] o snd)
                       |> map string_of_hyper_pconst));
           relevant [] [] hopeful
@@ -499,7 +509,7 @@
          |> insert_into_facts accepts
   in
     facts |> map_filter (pair_consts_fact thy is_built_in_const fudge)
-          |> iter 0 max_facts thres0 goal_const_tab []
+          |> iter 0 max_facts thres0 goal_const_tab goal_const_iter_tab []
           |> insert_special_facts
           |> tap (fn accepts => trace_msg ctxt (fn () =>
                       "Total relevant: " ^ string_of_int (length accepts)))
--- a/src/HOL/Tools/Sledgehammer/sledgehammer_run.ML	Fri Aug 23 20:53:00 2013 +0200
+++ b/src/HOL/Tools/Sledgehammer/sledgehammer_run.ML	Fri Aug 23 21:59:29 2013 +0200
@@ -167,8 +167,6 @@
 
 val auto_try_max_facts_divisor = 2 (* FUDGE *)
 
-fun eq_facts p = eq_list (op = o pairself fst) p
-
 fun string_of_facts facts =
   "Including " ^ string_of_int (length facts) ^
   " relevant fact" ^ plural_s (length facts) ^ ":\n" ^