lemur

This module performs learning over Logic Programs with Annotated Disjunctions and CP-Logic programs using the LEMUR algorithm of

Nicola Di Mauro, Elena Bellodi, and Fabrizio Riguzzi. Bandit-based Monte-Carlo structure learning of probabilistic logic programs. Machine Learning, 100(1):127-156, July 2015

See https://friguzzi.github.io/cplint/ for details.

Reexports slipcover

author: - Nicola di Mauro, Fabrizio Riguzzi, Elena Bellodi
copyright: - Nicola di Mauro, Fabrizio Riguzzi, Elena Bellodi */
license: - Artistic License 2.0

24:-reexport(library(slipcover)). 25 26:-use_module(library(lists)). 27:-use_module(library(random)). 28:-use_module(library(system)). 29:-use_module(library(terms)). 30 31 32:- set_prolog_flag(discontiguous_warnings,on). 33:- set_prolog_flag(single_var_warnings,on). 34:- set_prolog_flag(unknown,warning). 35 36:- dynamic db/1. 37:- dynamic lm_input_mod/1. 38 39 40:- meta_predicate induce_lm(:,-). 41:- meta_predicate induce_rules(:,-). 42:- meta_predicate set_lm(:,+). 43:- meta_predicate setting_lm(:,-). 44/* 45 declarations start 46define the depth of a variable appearing in a clause A B ^ : : : ^ Br as follows. 47 48Variables appearing in the head of a clause have depth zero. 49Otherwise, let Bi be the first literal containing the variable V, and let d be the maximal depth of the input variables of Bi 50then the depth of V is d + 1. The depth of a clause is the maximal depth of any variable in the clause. 51 52In questo modo possiamo lasciare il numero massimo di variabili a 4 (e cosi' impara la regola con taughtby) e riduciamo la profondita' massima delle variabili a 2 (in questo modo dovremmo imparare la regola con i due publication nel body e anche quella con taughtby). 53Bisogna modificare revise.pl per controllare che gli atomi che si aggiungono nel body non abbiano variabili oltre la profondita' massima. 54*/ 55 56 57default_setting_lm(epsilon_em,0.0001). 58default_setting_lm(epsilon_em_fraction,0.00001). 59default_setting_lm(eps,0.0001). 60default_setting_lm(eps_f,0.00001). 61 62/* if the difference in log likelihood in two successive em iteration is smaller 63than epsilon_em, then EM stops */ 64default_setting_lm(epsilon_sem,2). 65 66/* number of random restarts of em */ 67default_setting_lm(random_restarts_REFnumber,1). 68default_setting_lm(random_restarts_number,1). 69default_setting_lm(iterREF,-1). 70default_setting_lm(iter,-1). 71default_setting_lm(examples,atoms). 72default_setting_lm(group,1). 73default_setting_lm(d,1). 74default_setting_lm(verbosity,1). 75default_setting_lm(logzero,log(0.000001)). 76default_setting_lm(max_iter,10). 77default_setting_lm(max_iter_structure,10000). 78default_setting_lm(maxdepth_var,2). 79default_setting_lm(beamsize,100). 80default_setting_lm(background_clauses,50). 81default_setting_lm(neg_ex,cw). 82 83 84default_setting_lm(seed,seed(3032)). 85default_setting_lm(c_seed,21344). 86default_setting_lm(score,ll). 87/* allowed values: ll aucpr */ 88 89default_setting_lm(mcts_beamsize,3). 90default_setting_lm(mcts_visits,1e20). 91default_setting_lm(max_var,4). 92default_setting_lm(mcts_max_depth,8). 93default_setting_lm(mcts_c,0.7). 94default_setting_lm(mcts_iter,20). 95default_setting_lm(mcts_maxrestarts,20). 96default_setting_lm(mcts_covering,true). 97default_setting_lm(max_rules,1). 98default_setting_lm(epsilon_parsing, 1e-5). 99default_setting_lm(bagof,false). 100default_setting_lm(compiling,off). 101default_setting_lm(depth_bound,false). %if true, it limits the derivation of the example to the value of 'depth' 102default_setting_lm(depth,2). 103default_setting_lm(single_var,false). %false:1 variable for every grounding of a rule; true: 1 variable for rule (even if a rule has more groundings),simpler. 104default_setting_lm(tabling,auto). 105/* values: 106 auto 107 explicit 108*/ 109default_setting_lm(alpha,0.0). 110% Sets the type of parameter initialization for EM on Environment: 111% if alpha is 0.0, it uses a truncated Dirichlet process 112% if alpha is a float > 0.0, it uses a symmetric Dirichlet distribution 113% with that value as parameter 114:- thread_local database/1, lm_input_mod/1.

123induce_lm(M:TrainFolds,P):- 124 must_be(list,TrainFolds), 125 must_be(var,P), 126 induce_rules(M:TrainFolds,P0), 127 rules2terms(P0,P). 128 129 130 131induce_rules(M:Folds,R):- 132 set_lm(M:compiling,on), 133 M:local_setting(seed,Seed), 134 set_random(Seed), 135 M:local_setting(c_seed,CSeed), 136 rand_seed(CSeed), 137 findall(Exs,(member(F,Folds),M:fold(F,Exs)),L), 138 append(L,DB), 139 assert(M:database(DB)), 140 141 statistics(walltime,[_,_]), 142 format2(M,"\nMonte Carlo Tree Search for LPAD Structure Learning\n",[]), 143 144 (M:bg(RBG0)-> 145 process_clauses(RBG0,M,[],_,[],RBG), 146 generate_clauses(RBG,M,_RBG1,0,[],ThBG), 147 generate_clauses_bg(RBG,ClBG), 148 assert_all(ThBG,M,ThBGRef), 149 assert_all(ClBG,M,ClBGRef) 150 ; 151 true 152 ), 153 findall(BL , M:modeb(_,BL), BLS0), 154 sort(BLS0,BSL), 155 assert(M:mcts_modeb(BSL)), 156 157 assert(M:mcts_restart(1)), 158 learn_struct_mcts(DB,M,[],R2,Score2), 159 retract(M:mcts_restart(_)), 160 learn_params(DB,M,R2,R,Score), 161 162 format2(M,"~nRefinement score ~f - score after EMBLEM ~f~n",[Score2,Score]), 163 statistics(walltime,[_,WT]), 164 WTS is WT/1000, 165 write2(M,'\n\n'), 166 format2(M,'Wall time ~f */~n',[WTS]), 167 write_rules2(M,R,user_output), 168 set_lm(M:compiling,off), 169 (M:bg(RBG0)-> 170 retract_all(ThBGRef), 171 retract_all(ClBGRef) 172 ; 173 true 174 ). 175 176 177learn_struct_mcts(DB,M,R1,R,CLL1):- 178 learn_params(DB,M, R1, R3, CLL), 179 write2(M,'updated Theory'),nl2(M), 180 write_rules2(M,R3,user_output), 181 182 assert(M:mcts_best_score(CLL)), 183 assert(M:mcts_best_theory(R3)), 184 assert(M:mcts_theories(0)), 185 186 assert(M:mcts_best_theories_iteration([])), 187 188 mcts(R3,M,CLL,DB), 189 retract(M:mcts_best_theories_iteration(BestsIter)), 190 format2(M,"\nBests found at : ~w",[BestsIter]), 191 192 retract(M:mcts_theories(_)), 193 retract(M:mcts_best_score(CLLNew)), 194 retract(M:mcts_best_theory(RNew)), 195 196 (M:local_setting(mcts_covering,true) -> 197 198 M:local_setting(mcts_maxrestarts,MctsRestarts), 199 M:mcts_restart(CurrentRestart), 200 201 Improvement is CLLNew - CLL, 202 ( (CLLNew > CLL, Improvement > 0.1, CurrentRestart =< MctsRestarts) -> 203 204 format2(M,"\n---------------- Improvement ~w",[Improvement]), 205 retractall(M:node(_, _, _, _, _, _, _)), 206 retract(M:local_setting(max_rules,ParRules)), 207 ParRules1 is ParRules + 1, 208 assert(M:local_setting(max_rules,ParRules1)), 209 retract(M:mcts_restart(Restart)), 210 Restart1 is Restart + 1, 211 assert(M:mcts_restart(Restart1)), 212 learn_struct_mcts(DB,M,RNew,R,CLL1) 213 ; 214 CLL1 = CLLNew, 215 R = RNew 216 ) 217 ; 218 CLL1 = CLLNew, 219 R = RNew 220 ). 221 222 223 224mcts(InitialTheory,M,InitialScore,DB):- 225 % node(ID, CHILDRENS, PARENT, CLL, Theory, VISITED, BACKSCORE) 226 assert(M:node(1, [], 0, InitialScore , InitialTheory, 0 , 0)), 227 assert(M:lastid(1)), 228 M:local_setting(mcts_iter,I), 229 assert(M:mcts_iteration(0)), 230 cycle_mcts(I,M,DB), 231 retract(M:mcts_iteration(_)), 232 retract(M:lastid(Nodes)), 233 %print_graph, 234 format2(M,"\nTree size: ~w nodes.",[Nodes]). 235 236 237backup_amaf(1,M,_Reward,_):- 238 !, 239 (retract(M:node(1, Childs, Parent , PSLL, MLN, Visited, Backscore)) -> 240 true 241 ; 242 format(user_error,"\nNo node with ID ~w in backup",[NodeID]), 243 throw(no_node_id(NodeID)) 244 ), 245 Visited1 is Visited + 1, 246 assert(M:node(1, Childs, Parent , PSLL, MLN, Visited1, Backscore)). 247 248backup_amaf(NodeID,M,Reward,ParentsTranspose):- 249 (retract(M:node(NodeID, Childs, Parent , PSLL, MLN, Visited, Backscore)) -> 250 true 251 ; 252 format(user_error,"\nNo node with ID ~w in backup",[NodeID]), 253 throw(no_node_id(NodeID)) 254 ), 255 (member(NodeID,ParentsTranspose) -> 256 Backscore1 is Backscore, 257 Visited1 is Visited 258 %format("~w- ",[NodeID]) 259 ; 260 (PSLL =:= 1 -> 261 Backscore1 is Backscore + Reward 262 ; 263 SigmoidValue is 1 / (1 - PSLL), 264 ( Reward > SigmoidValue -> 265 Backscore1 is Backscore + Reward 266 ; 267 Backscore1 is Backscore + SigmoidValue 268 %Backscore1 is Backscore + Reward 269 ) 270 ), 271 272 Visited1 is Visited + 1 273 %format("~w+ ",[NodeID]) 274 ), 275 assert(M:node(NodeID, Childs, Parent , PSLL, MLN, Visited1, Backscore1)). 276 277 278check_amaf(NodeID,M,Theory,SigmoidValue,ParentsTranspose):- 279 M:lastid(Nodes), 280 format2(M,"\nChecking amaf: node ~w, parents ~w: ",[NodeID,ParentsTranspose]), 281 check_amaf(Nodes,M,NodeID,Theory,SigmoidValue,ParentsTranspose). 282 283check_amaf(1,M,_NodeID,_,_SigmoidValue,_ParentsTranspose):- 284 retract(M:node(1, Childs, Parent , PSLL, MLN, Visited, Backscore)), 285 Visited1 is Visited + 1, 286 assert(M:node(1, Childs, Parent , PSLL, MLN, Visited1, Backscore)), 287 !. 288 289check_amaf(Node,M,NodeID,Theory,SigmoidValue,ParentsTranspose):- 290 Node \== NodeID, 291 !, 292 M:node(Node, _Childs, _Parent , _CLL, TheoryN, _Visited, _Backscore), 293 ( subsume_theory(TheoryN,Theory) -> 294 %format("\n\t ~w ~w: ",[TheoryN,Theory]), 295 backup_amaf(Node,M,SigmoidValue,ParentsTranspose) 296 ; 297 true 298 ), 299 Node1 is Node - 1, 300 check_amaf(Node1,M,NodeID,Theory,SigmoidValue,ParentsTranspose). 301 302check_amaf(Node,M,NodeID,Theory,SigmoidValue,ParentsTranspose):- 303 Node1 is Node - 1, 304 check_amaf(Node1,M,NodeID,Theory,SigmoidValue,ParentsTranspose). 305 306 307subsume_theory(Theory,TheoryN):- 308 copy_term(Theory,Theory1), 309 skolemize(TheoryN,TheoryN1), 310 subsume_theory1(Theory1,TheoryN1), 311 !. 312 313skolemize(Theory,Theory1):- 314 copy_term(Theory,Theory1), 315 term_variables(Theory1,Vars), 316 skolemize1(Vars,1). 317 318 319skolemize1([],_). 320 321skolemize1([Var|R],K):- 322 atomic_list_concat([s,K],Skolem), 323 Var = Skolem, 324 K1 is K + 1, 325 skolemize1(R,K1). 326 327subsume_theory1([],_). 328 329subsume_theory1([Rule|R],TheoryN):- 330 subsume_theory2(Rule,TheoryN,NewTheoryN), 331 subsume_theory1(R,NewTheoryN). 332 333 334subsume_theory2(Rule,[Rule1|R],R):- 335 Rule = rule(_,[H: _, _: _],Body,_,_), 336 Rule1 = rule(_,[H1: _, _: _],Body1,_,_), 337 H = H1, 338 subsume_body(Body,Body1), 339 !. 340 341subsume_theory2(Rule,[Rule1|R],[Rule1|R1]):- 342 subsume_theory2(Rule,R,R1). 343 344 345subsume_body(Body,Body1):- 346 length(Body,L), 347 length(Body1,L1), 348 L =< L1, 349 subsume_body1(Body,Body1). 350 351 352subsume_body1([],_). 353 354subsume_body1([L|R],Body):- 355 nth1(_,Body,L,Rest), 356 subsume_body1(R,Rest). 357 358 359same_theory(Theory0,TheoryN):- 360 copy_term(Theory0,Theory), 361 length(Theory,L), 362 length(TheoryN,L), 363 same_theory1(Theory,TheoryN), 364 !. 365 366 367same_theory1([],[]). 368 369same_theory1([Rule|R],TheoryN):- 370 same_theory2(Rule,TheoryN,NewTheoryN), 371 same_theory1(R,NewTheoryN). 372 373 374same_theory2(Rule,[Rule1|R],R):- 375 Rule = rule(_,[H: _, _: _],Body,_,_), 376 Rule1 = rule(_,[H1: _, _: _],Body1,_,_), 377 H = H1, 378 same_body(Body,Body1), 379 !. 380 381same_theory2(Rule,[Rule1|R],[Rule1|R1]):- 382 same_theory2(Rule,R,R1). 383 384 385same_body(Body,Body1):- 386 length(Body,L), 387 length(Body1,L), 388 same_body1(Body,Body1). 389 390 391same_body1([],[]). 392 393same_body1([L|R],Body):- 394 nth1(_,Body,L,Rest), 395 same_body1(R,Rest). 396 397 398cycle_mcts(0,_M,_):- 399 !. 400 401cycle_mcts(K,M,DB):- 402 M:local_setting(mcts_iter,MaxI), 403 Iteration is MaxI - K + 1, 404 retract(M:mcts_iteration(_)), 405 assert(M:mcts_iteration(Iteration)), 406 format2(M,"\nIteration ~w",[Iteration]), 407 tree_policy(1,M,NodeID,DB,1,_Depth), 408 ( M:node(NodeID, _Childs, _Parent , _CLL, Theory, _Visited, _Backscore) -> 409 % do update with the sigmoid of the Score 410 % SigmoidValue is ((1 / (1 + exp(-PSLL)))/0.5), 411 % format("\n~w: ~w ~w Sigmoid ~w",[K,MLN,PSLL,SigmoidValue]), 412 M:local_setting(mcts_max_depth, MaxDepth), 413 random(1,MaxDepth,MaxDepth1), 414 default_policy(Theory,M,-1e20,Reward,_,BestDefaultTheory,DB,1,MaxDepth1), 415 % do update with the sigmoid of the Score 416 % SigmoidValue is ((1 / (1 + exp(-Reward)))/0.5), 417 418 (Reward=:=1-> 419 SigmoidValue=1e20 420 ; 421 SigmoidValue is 1 / (1 - Reward) 422 ), 423 ( Reward =\= -1e20 -> 424 425 % (Reward > CLL -> 426 % SigmoidValue = 1 427 % ; 428 % SigmoidValue = 0 429 % ), 430 431 % format("\n~w: ~w \nReward ~w Sigmoid ~w",[K,Theory,Reward,SigmoidValue]), 432 format2(M,"\n[Backup reward ~w]",[SigmoidValue]), 433 backup(NodeID,M,SigmoidValue,Parents), 434 % check_transposition(NodeID,Theory,SigmoidValue,Parents), 435 check_amaf(NodeID,M,BestDefaultTheory,SigmoidValue,Parents) 436 ; 437 format2(M,"\n--> no default policy expansion",[]) 438 ), 439 K1 is K - 1, 440 %read(_), 441 cycle_mcts(K1,M,DB) 442 ; 443 format2(M,"\n--> tree policy end",[]) 444 ). 445 446 447prune([],_Childs1). 448 449prune([ID|R],Childs1):- 450 member(ID,Childs1), 451 !, 452 prune(R,Childs1). 453 454prune([ID|R],Childs1):- 455 prune_sub_tree(ID), 456 prune(R,Childs1). 457 458 459prune_sub_tree(ID):- 460 retract(node(ID, Childs, _Parent , _CLL, _Theory, _VISITED, _BACKSCORE)), 461 prune_sub_tree1(Childs). 462 463 464prune_sub_tree1([]). 465 466prune_sub_tree1([ID|R]):- 467 retract(node(ID, Childs, _Parent , _CLL, _Theory, _VISITED, _BACKSCORE)), 468 prune_sub_tree1(Childs), 469 prune_sub_tree1(R). 470 471 472tree_policy(ID,M,NodeID,DB,Od,Nd):- 473 %check_pruning(ID), 474 475 (retract(M:node(ID, Childs, Parent , CLL, Theory, VISITED, BACKSCORE)) -> 476 true 477 ; 478 throw(no_node_id(ID)) 479 ), 480 %format("\n Tree policy ~w ~w ~w",[Theory,VISITED, BACKSCORE]), 481 format2(M,"\n[Tree Policy ~w, ~w, ~w] ",[ID,VISITED,BACKSCORE]), flush_output, 482 %( VISITED = 0, ID \= 1 -> 483 ( CLL = 1, ID \= 1 -> 484 score_theory(Theory,M,DB,CLL1,BestTheory,NewTheory), 485 M:mcts_best_score(BestScore), 486 487 ( M:local_setting(mcts_covering,true) -> 488 length(NewTheory,NewTheoryL), %lemurc 489 length(Theory,TheoryL), 490 ( NewTheoryL = TheoryL -> 491 LengthCondition = true 492 ; 493 LengthCondition = false 494 ) 495 ; 496 LengthCondition = true 497 ), 498 499 500 ( ( CLL1 > BestScore, LengthCondition = true) -> 501 format2(M,"\n[New best score: ~w ~w]",[CLL1, BestTheory]),flush_output, 502 503 504 retract(M:mcts_best_score(_)), 505 retract(M:mcts_best_theory(_)), 506 assert(M:mcts_best_score(CLL1)), 507 assert(M:mcts_best_theory(NewTheory)), 508 509 retract(M:mcts_best_theories_iteration(BestsIter)), 510 M:mcts_iteration(Iteration), 511 append(BestsIter,[Iteration],BestsIter1), 512 assert(M:mcts_best_theories_iteration(BestsIter1)), 513 514 retract(M:mcts_theories(Mlns)), 515 Mlns1 is Mlns + 1, 516 assert(M:mcts_theories(Mlns1)) 517 ; 518 true 519 ) 520 ; 521 CLL1 = CLL, 522 NewTheory = Theory 523 ), 524 525 Visited1 is VISITED + 1, 526 527 % (CLL = 1 -> 528 % Visited2 = Visited1, 529 % (Visited2 == 2 -> Backscore1 = BACKSCORE ; Backscore1 = 0) % in this case the node has been visited by transposition 530 % ; 531 % Visited2 = Visited1, 532 % Backscore1 = BACKSCORE 533 % ), 534 535 Visited2 = Visited1, 536 Backscore1 = BACKSCORE, 537 538 539 (Childs == [] -> 540 Nd = Od, 541 expand(ID,M, Theory, CLL1, DB, NodeID, Childs1), 542 assert(M:node(ID, Childs1, Parent , CLL1, NewTheory, Visited2, Backscore1)) 543 ; 544 Od1 is Od + 1, 545 minmaxvalue(Childs,M,MinV,MaxV), 546 %mean_value_level(Childs,M,Mvl), 547 once(uct(Childs, M,VISITED, MinV, MaxV, BestChild)), 548 %once(uct(Childs, VISITED, BestChild)), 549 tree_policy(BestChild,M,NodeID,DB,Od1, Nd), 550 assert(M:node(ID, Childs, Parent , CLL1, NewTheory, Visited2, Backscore1)) 551 ). 552 553 554default_policy(_Theory, _M,Reward, Reward, BestDefaultTheory,BestDefaultTheory,_DB, Depth, MaxDepth):- 555 Depth > MaxDepth, 556 !. 557 558default_policy(Theory,M,PrevR,Reward,PrevBestDefaultTheory,BestDefaultTheory,DB,Depth,MaxDepth):- 559 %format("\nDefault policy",[]),flush_output, 560 format2(M,"\n[Default Policy ~w]",[Depth]), 561 theory_revisions_r(Theory,M,Revisions), 562 ( Revisions \== [] -> 563 length(Revisions,L), 564 random(0,L,K), 565 nth0(K, Revisions,Spec), 566 Depth1 is Depth + 1, 567 score_theory(Spec,M,DB,Score,BestTheory,NewTheory), 568 ( M:local_setting(mcts_covering,true) -> 569 length(NewTheory,NewTheoryL), %lemurc 570 length(Spec,TheoryL), 571 ( NewTheoryL = TheoryL -> 572 LengthCondition = true 573 ; 574 LengthCondition = false 575 ) 576 ; 577 LengthCondition = true 578 ), 579 580 581 (( Score > PrevR, LengthCondition = true) -> 582 Reward1 = Score, 583 BestDefaultTheory1 = NewTheory 584 ; 585 Reward1 = PrevR, 586 BestDefaultTheory1 = PrevBestDefaultTheory 587 ), 588 589 format2(M," cll-reward ~w",[Reward1]), 590 591 M:mcts_best_score(BestScore), 592 593 594 ((Score > BestScore, LengthCondition = true) -> 595 format2(M,"\n[New best score: ~w ~w]",[Score, BestTheory]),flush_output, 596 597 598 retract(M:mcts_best_score(_)), 599 retract(M:mcts_best_theory(_)), 600 assert(M:mcts_best_score(Score)), 601 assert(M:mcts_best_theory(NewTheory)), 602 603 retract(M:mcts_best_theories_iteration(BestsIter)), 604 M:mcts_iteration(Iteration), 605 append(BestsIter,[Iteration],BestsIter1), 606 assert(M:mcts_best_theories_iteration(BestsIter1)), 607 608 609 retract(M:mcts_theories(Mlns)), 610 Mlns1 is Mlns + 1, 611 assert(M:mcts_theories(Mlns1)) 612 ; 613 true 614 ), 615 616 default_policy(Spec,M, Reward1,Reward, BestDefaultTheory1,BestDefaultTheory,DB, Depth1,MaxDepth) 617 618 ; 619 Reward = PrevR, 620 BestDefaultTheory = PrevBestDefaultTheory 621 ). 622 623 624minmaxvalue(Childs,M,MinV,MaxV):- 625 Childs = [F|R], 626 M:node(F, _, _ , _, _, Visits, Reward), 627 ( Visits=:=0-> 628 V is sign(Reward)*1e20 629 ; 630 V is Reward / Visits 631 ), 632 minmaxvalue(R,M,V,V,MinV,MaxV). 633 634minmaxvalue([],_M,Min,Max,Min,Max). 635 636minmaxvalue([C|R],M,PrevMin,PrevMax,MinV,MaxV):- 637 M:node(C, _, _ , _, _, Visits, Reward), 638 ( Visits=:=0-> 639 V is sign(Reward)*1e20 640 ; 641 V is Reward / Visits 642 ), 643 ( V > PrevMax -> 644 Max1 is V 645 ; 646 Max1 is PrevMax 647 ), 648 ( V < PrevMin -> 649 Min1 is V 650 ; 651 Min1 is PrevMin 652 ), 653 minmaxvalue(R,M,Min1,Max1,MinV,MaxV). 654 655 656mean_value_level(Cs,Mod,M):- 657 mean_value_level1(Cs,Mod,Me), 658 length(Me,L), 659 sum_list(Me,S), 660 ( L=:=0-> 661 M is sign(S)*1e20 662 ; 663 M is S / L 664 ). 665 666 667mean_value_level1([],_Mod,[]). 668 669mean_value_level1([C|R],Mod,M1):- 670 Mod:node(C, _, _ , 1, _, _Visits, _Reward), 671 !, 672 mean_value_level1(R,Mod,M1). 673 674mean_value_level1([C|R],Mod,[M|Rm]):- 675 Mod:node(C, _, _ , _, _, Visits, Reward), 676 !, 677 mean_value_level1(R,Mod,Rm), 678 ( Visits=:=0-> 679 M is sign(Reward)*1e20 680 ; 681 M is (Reward / Visits) 682 ). 683 684 685uct(Childs, M,ParentVisits, Min, Max, BestChild):- 686 Childs = [FirstChild|RestChilds], 687 M:node(FirstChild, _, _ , _Score, _Theory, Visits, Reward), 688 ( Visits == 0 -> 689 BestChild = FirstChild 690 ; 691 M:local_setting(mcts_c,C), 692% ( Score == 1 -> 693% R is Mvl 694% ; 695% R is Reward 696% ), 697 ( Max-Min=:=0-> 698 UCT is sign(Reward/Visits-Min)*1e20 699 ; 700 R is Reward, 701 %AA is ((R / Visits) - Min ) / (Max-Min), 702 %BB is 2 * C * sqrt(2 * log(ParentVisits) / Visits), 703 UCT is ((R / Visits) - Min ) / (Max-Min) + 2 * C * sqrt(2 * log(ParentVisits) / Visits) 704 ), 705%%% format("\nID ~w UCT ~w ~w/~w=~w ~w",[FirstChild,UCT,R,Visits,AA,BB]), 706%%% format("\n\t ~w ~w",[Score,Theory]), 707%%% format("~w ",[UCT]), 708 uct(RestChilds,M, UCT, ParentVisits, FirstChild, Min,Max, BestChild) 709 ). 710 711 712uct([],_M, _CurrentBestUCT, _ParentVisits, BestChild, _, _,BestChild). 713 714uct([Child|RestChilds], M,CurrentBestUCT, ParentVisits, CurrentBestChild, Min, Max,BestChild) :- 715 M:node(Child, _, _ , _Score, _Theory, Visits, Reward), 716 ( Visits == 0 -> 717 BestChild = Child 718 ; 719 M:local_setting(mcts_c,C), 720% ( Score == 1 -> 721% R is Mvl 722% ; 723% R is Reward 724% ), 725 ( Max-Min=:=0-> 726 UCT is sign(Reward/Visits-Min)*1e20 727 ; 728 R is Reward, 729 %AA is ((R / Visits) - Min ) / (Max-Min), 730 %BB is 2 * C * sqrt(2 * log(ParentVisits) / Visits), 731 UCT is ((R / Visits) - Min ) / (Max-Min) + 2 * C * sqrt(2 * log(ParentVisits) / Visits) 732 ), 733%%% format("\nID ~w UCT ~w ~w/~w=~w ~w",[Child,UCT,R,Visits,AA,BB]), 734%%% format("\n\t ~w ~w",[Score,Theory]), 735%%% format("~w ",[UCT]),flush_output, 736 ( UCT > CurrentBestUCT -> 737 uct(RestChilds,M, UCT, ParentVisits, Child, Min, Max, BestChild) 738 ; 739 uct(RestChilds,M, CurrentBestUCT, ParentVisits, CurrentBestChild, Min, Max, BestChild) 740 ) 741 ). 742 743 744expand(ID, M,Theory, ParentCLL, DB, NodeID, Childs):- 745 %format(" expanding...",[]),flush_output, 746 theory_revisions(Theory,M,Revisions), 747 !, 748 assert_childs(Revisions,M,ID,ParentCLL,Childs), 749 ( Childs \= [] -> 750 Childs = [NodeID|_], 751 retract(M:node(NodeID, Childs1, Parent , _, Theory1, Visited, Backscore)), 752 format2(M,"\n[Expand ~w]",[NodeID]), 753 Visited1 is Visited + 1, 754 score_theory(Theory1,M,DB,CLL,BestTheory,NewTheory), 755 format2(M," CLL: ~w]",[CLL]), 756 %format("\nTree policy: ~w ~w]",[Score, Theory1]), 757 M:mcts_best_score(BestScore), 758 759 %Ratio is BestScore / CLL, 760 %( Ratio > 1.001 -> 761 762 ( M:local_setting(mcts_covering,true) -> 763 length(NewTheory,NewTheoryL), %lemurc 764 length(Theory1,Theory1L), 765 ( NewTheoryL = Theory1L -> 766 LengthCondition = true 767 ; 768 LengthCondition = false 769 ) 770 ; 771 LengthCondition = true 772 ), 773 774 775 ( ( CLL > BestScore, LengthCondition = true) -> 776 format2(M,"\n[New best score: ~w ~w]",[CLL, BestTheory]),flush_output, 777 retract(M:mcts_best_score(_)), 778 retract(M:mcts_best_theory(_)), 779 assert(M:mcts_best_score(CLL)), 780 assert(M:mcts_best_theory(NewTheory)), 781 782 retract(M:mcts_best_theories_iteration(BestsIter)), 783 M:mcts_iteration(Iteration), 784 append(BestsIter,[Iteration],BestsIter1), 785 assert(M:mcts_best_theories_iteration(BestsIter1)), 786 787 788 retract(M:mcts_theories(Mlns)), 789 Mlns1 is Mlns + 1, 790 assert(M:mcts_theories(Mlns1)) 791 ; 792 true 793 ), 794 assert(M:node(NodeID, Childs1, Parent , CLL, NewTheory, Visited1, Backscore)) 795 ; 796 NodeID = -1 797 ). 798 799 800assert_childs([],_M,_,_,[]). 801 802assert_childs([Spec|Rest],M,P,PCLL,[ID1|Childs]):- 803 %node(ID, CHILDRENS, PARENT, PSLL, MLN, VISITED, BACKSCORE) 804 retract(M:lastid(ID)), 805 %format(" ~w",[ID]),flush_output, 806 ID1 is ID + 1, 807 assert(M:lastid(ID1)), 808 %SigmoidValue is ((1 / (1 + exp(-PCLL)))/0.5), 809 (PCLL=:=1-> 810 SigmoidValue=1e20 811 ; 812 SigmoidValue is 1 / (1 - PCLL) 813 ), 814 assert(M:node(ID1, [], P, 1 , Spec, 1 , SigmoidValue)), 815 %assert(node(ID1, [], P, 1 , Spec, 0 , 0)), 816 assert_childs(Rest,M,P,PCLL,Childs). 817 818 819theory_length([],X,X). 820 821theory_length([T|R],K,K1):- 822 theory_length(R,K,K0), 823 T = rule(_,_,B,_,_), 824 length(B,L), 825 ( L > K0 -> 826 K1 = L 827 ; 828 K1 = K0 829 ). 830 831score_theory(Theory0,M,DB,Score,Theory,R3):- 832 ( M:mcts_theories(0) -> 833 Theory = Theory0 834 ; 835 theory_length(Theory0,0,Le), 836 ( Le > 1 -> 837% mcts_best_theory(TheoryBest), 838% append(TheoryBest,Theory0,Theory) 839 Theory = Theory0 840 ; 841 Theory = Theory0 842 ) 843 ), 844 learn_params(DB, M, Theory, R3, CLL), 845 write3(M,'Updated refinement'),write3(M,'\n'), 846 write_rules3(M,R3,user_output), 847 Score = CLL, 848 !. 849 850 851backup(1,_M,_Reward,[]):- 852 !. 853 854backup(NodeID,M,Reward,[Parent|R]):- 855 ( retract(M:node(NodeID, Childs, Parent , PSLL, MLN, Visited, Backscore)) -> 856 true 857 ; 858 format2(M,"\nNo node with ID ~w in backup",[NodeID]), 859 throw(no_node_id(NodeID)) 860 ), 861 ( PSLL=:=1-> 862 SigmoidValue=1e20 863 ; 864 SigmoidValue is 1 / (1 - PSLL) 865 ), 866 ( Reward > SigmoidValue -> 867 Backscore1 is Backscore + Reward, 868 Reward1 is Reward 869 ; 870 Backscore1 is Backscore + SigmoidValue, 871 Reward1 is SigmoidValue 872 %Backscore1 is Backscore + Reward, 873 %Reward1 is Reward 874 ), 875 %format("\n backup ~w ~w",[NodeID,MLN]), 876 assert(M:node(NodeID, Childs, Parent , PSLL, MLN, Visited, Backscore1)), 877 backup(Parent,M,Reward1,R). 878 879 880theory_revisions_op(Theory,M,TheoryRevs):- 881 setof(RevOp, Theory^revise_theory(Theory,M,RevOp), TheoryRevs),!. 882 883theory_revisions_op(_Theory,_M,[]). 884 885 886theory_revisions_r(Theory,M,TheoryRevs):- 887 theory_revisions_op(Theory,M,TheoryRevs1), 888 %filter_add_rule(TheoryRevs11,TheoryRevs1), 889 890 ( TheoryRevs1 == [] -> 891 TheoryRevs = [] 892 ; 893 length(TheoryRevs1,L), 894 random(0,L,K), 895 nth0(K, TheoryRevs1,Revision), 896 apply_operators([Revision],Theory,TheoryRevs) 897 ). 898 899 900theory_revisions(Theory,M,TheoryRevs):- 901 theory_revisions_op(Theory,M,TheoryRevs1), 902 apply_operators(TheoryRevs1,Theory,TheoryRevs). 903 904 905apply_operators([],_Theory,[]). 906 907apply_operators([add(Rule)|RestOps],Theory,[NewTheory|RestTheory]) :-!, 908 append(Theory, [Rule], NewTheory), 909 %nl,write(NewTheory), 910 apply_operators(RestOps,Theory,RestTheory). 911 912apply_operators([add_body(Rule1,Rule2,_A)|RestOps],Theory,[NewTheory|RestTheory]) :-!, 913 delete_matching(Theory,Rule1,Theory1), 914 append(Theory1, [Rule2], NewTheory), 915 %nl,write(NewTheory), 916 apply_operators(RestOps,Theory,RestTheory). 917 918revise_theory(Theory,M,Ref):- 919 specialize_theory(Theory,M,Ref). 920 921revise_theory(Theory,M,Ref):- 922 generalize_theory(Theory,M,Ref). 923 924 925generalize_theory(Theory,M,Ref):- 926 length(Theory,LT), 927 M:local_setting(max_rules,MR), 928 LT<MR, 929 add_rule(M,Ref). 930 931 932add_rule(M,add(SpecRule)):- 933 findall(HL , M:modeh(_,HL), HLS), 934 length(HLS,L), 935 L1 is L+1, 936 P is 1/L1, 937 generate_head(HLS,P,Head), 938 get_next_rule_number(M,ID), 939 Rule0 = rule(ID,Head,[],true,_), 940 specialize_rule(Rule0,M,SpecRule,_Lit). 941 942generate_head([H|_T],_P,[H1:0.5,'':0.5]):- 943 H=..[Pred|Args], 944 length(Args,LA), 945 length(Args1,LA), 946 H1=..[Pred|Args1]. 947 948generate_head([_H|T],P,Head):- 949 generate_head(T,P,Head). 950 951 952specialize_theory(Theory,M,Ref):- 953 Theory \== [], 954 choose_rule(Theory,M,Rule), 955 specialize_rule(Rule,M,SpecRule,Lit), 956 Ref = add_body(Rule,SpecRule,Lit), 957 SpecRule = rule(_,_,_B,_,_). 958 959 960specialize_rule(Rule,M,SpecRule,Lit):- 961 M:mcts_modeb(BSL), 962 specialize_rule_bl(BSL,M,Rule,SpecRule,Lit). 963 964 965 966 967 968specialize_rule_bl([Lit|_RLit],M,Rule,SpecRul,SLit):- 969 Rule = rule(ID,LH,BL,true,Tun), 970 remove_prob(LH,LH1), 971 append(LH1,BL,ALL), 972 specialize_rule_lit(Lit,M,ALL,SLit), 973 append(BL,[SLit],BL1), 974 (M:lookahead(SLit,LLit1);M:lookahead_cons(SLit,LLit1)), 975 specialize_rule_la(LLit1,M,LH1,BL1,BL2), 976 append(LH1,BL2,ALL2), 977 extract_fancy_vars(ALL2,Vars1), 978 length(Vars1,NV), 979 M:local_setting(max_var,MV), 980 NV=<MV, 981 \+ banned_clause(M,LH1,BL2), 982 SpecRul = rule(ID,LH,BL2,true,Tun). 983 984specialize_rule_bl([Lit|_RLit],M,Rule,SpecRul,SLit):- 985 Rule = rule(ID,LH,BL,true,Tun), 986 remove_prob(LH,LH1), 987 append(LH1,BL,ALL), 988 specialize_rule_lit(Lit,M,ALL,SLit), 989 990 \+ M:lookahead_cons(SLit,_), 991 992 append(BL,[SLit],BL1), 993 append(LH1,BL1,ALL1), 994 extract_fancy_vars(ALL1,Vars1), 995 length(Vars1,NV), 996 M:local_setting(max_var,MV), 997 NV=<MV, 998 M:local_setting(maxdepth_var,_MD), 999 \+ banned_clause(M,LH1,BL1), 1000 SpecRul = rule(ID,LH,BL1,true,Tun). 1001 1002 1003specialize_rule_bl([_|RLit],M,Rule,SpecRul,Lit):- 1004 specialize_rule_bl(RLit,M,Rule,SpecRul,Lit). 1005 1006 1007specialize_rule_la([],_M,_LH1,BL1,BL1). 1008 1009specialize_rule_la([Lit1|T],M,LH1,BL1,BL3):- 1010 copy_term(Lit1,Lit2), 1011 M:modeb(_,Lit2), 1012 append(LH1,BL1,ALL1), 1013 specialize_rule_lit(Lit2,M,ALL1,SLit1), 1014 append(BL1,[SLit1],BL2), 1015 specialize_rule_la(T,M,LH1,BL2,BL3). 1016 1017 1018 1019remove_prob(['':_P],[]):-!. 1020 1021remove_prob([X:_|R],[X|R1]):- 1022 remove_prob(R,R1). 1023 1024 1025specialize_rule_lit(Lit,M,Lits,SpecLit):- 1026 Lit =.. [Pred|Args], 1027 extract_type_vars(Lits,M,TypeVars0), 1028 remove_duplicates(TypeVars0,TypeVars), 1029 take_var_args(Args,TypeVars,Args1), 1030 SpecLit =.. [Pred|Args1], 1031 \+ member_eq(SpecLit,Lits). 1032 1033choose_rule(Theory,M,Rule):- 1034 ( M:local_setting(mcts_covering,true) -> 1035 M:mcts_restart(Restart), 1036 nth1(K,Theory,Rule), 1037 K >= Restart 1038 ; 1039 member(Rule,Theory) 1040 ). 1041 1042 1043 1044delete_matching([],_El,[]). 1045 1046delete_matching([El|T],El,T1):-!, 1047 delete_matching(T,El,T1). 1048 1049delete_matching([H|T],El,[H|T1]):- 1050 delete_matching(T,El,T1).

1073setting_lm(M:P,V):- 1074 must_be(atom,P), 1075 M:local_setting(P,V). 1076 1077 1078 1079 1080 1081 1082 1083 1084lemur_expansion((:- begin_bg), []) :- 1085 prolog_load_context(module, M), 1086 lm_input_mod(M),!, 1087 assert(M:bg_on). 1088 1089lemur_expansion(C, M:bgc(C)) :- 1090 prolog_load_context(module, M), 1091 C\= (:- end_bg), 1092 lm_input_mod(M), 1093 M:bg_on,!. 1094 1095lemur_expansion((:- end_bg), []) :- 1096 prolog_load_context(module, M), 1097 lm_input_mod(M),!, 1098 retractall(M:bg_on), 1099 findall(C,M:bgc(C),L), 1100 retractall(M:bgc(_)), 1101 (M:bg(BG0)-> 1102 retract(M:bg(BG0)), 1103 append(BG0,L,BG), 1104 assert(M:bg(BG)) 1105 ; 1106 assert(M:bg(L)) 1107 ). 1108 1109lemur_expansion((:- begin_in), []) :- 1110 prolog_load_context(module, M), 1111 lm_input_mod(M),!, 1112 assert(M:in_on). 1113 1114lemur_expansion(C, M:inc(C)) :- 1115 prolog_load_context(module, M), 1116 C\= (:- end_in), 1117 lm_input_mod(M), 1118 M:in_on,!. 1119 1120lemur_expansion((:- end_in), []) :- 1121 prolog_load_context(module, M), 1122 lm_input_mod(M),!, 1123 retractall(M:in_on), 1124 findall(C,M:inc(C),L), 1125 retractall(M:inc(_)), 1126 (M:in(IN0)-> 1127 retract(M:in(IN0)), 1128 append(IN0,L,IN), 1129 assert(M:in(IN)) 1130 ; 1131 assert(M:in(L)) 1132 ). 1133 1134lemur_expansion(output(P/A), [output(P/A)|TabDir]) :- 1135 prolog_load_context(module, M), 1136 lm_input_mod(M), 1137 M:local_setting(tabling,auto),!, 1138 tab(M,P/A,P1), 1139 zero_clause(M,P/A,Z), 1140 term_expansion((:- table P1),TabDir), 1141 assert(M:zero_clauses([Z])). 1142 1143lemur_expansion(input(P/A), [input(P/A)|TabDir]) :- 1144 prolog_load_context(module, M), 1145 lm_input_mod(M), 1146 M:local_setting(tabling,auto),!, 1147 tab(M,P/A,P1), 1148 zero_clause(M,P/A,Z), 1149 term_expansion((:- table P1),TabDir), 1150 assert(M:zero_clauses([Z])). 1151 1152lemur_expansion(begin(model(I)), []) :- 1153 prolog_load_context(module, M), 1154 lm_input_mod(M),!, 1155 retractall(M:model(_)), 1156 assert(M:model(I)), 1157 assert(M:int(I)). 1158 1159lemur_expansion(end(model(_I)), []) :- 1160 prolog_load_context(module, M), 1161 lm_input_mod(M),!, 1162 retractall(M:model(_)). 1163 1164lemur_expansion(At, A) :- 1165 prolog_load_context(module, M), 1166 lm_input_mod(M), 1167 M:model(Name), 1168 At \= (_ :- _), 1169 At \= end_of_file, 1170 (At=neg(Atom)-> 1171 Atom=..[Pred|Args], 1172 Atom1=..[Pred,Name|Args], 1173 A=neg(Atom1) 1174 ; 1175 (At=prob(Pr)-> 1176 A='$prob'(Name,Pr) 1177 ; 1178 At=..[Pred|Args], 1179 Atom1=..[Pred,Name|Args], 1180 A=Atom1 1181 ) 1182 ). 1183 1184 1185:- multifile sandbox:safe_meta/2. 1186 1187sandbox:safe_meta(lemur:induce_lm(_,_),[]). 1188sandbox:safe_meta(lemur:set_lm(_,_), []). 1189sandbox:safe_meta(lemur:setting_lm(_,_), []). 1190 1191:- thread_local lemur_file/1. 1192 1193 1194user:term_expansion((:- lemur), []) :-!, 1195 prolog_load_context(source, Source), 1196 asserta(lemur_file(Source)), 1197 prolog_load_context(module, M), 1198 retractall(M:local_setting(_,_)), 1199 findall(local_setting(P,V),default_setting_lm(P,V),L), 1200 assert_all(L,M,_), 1201 assert(lm_input_mod(M)), 1202 retractall(M:rule_sc_n(_)), 1203 assert(M:rule_sc_n(0)), 1204 retractall(M:rule_ng_sc_n(_)), 1205 assert(M:rule_ng_sc_n(0)), 1206 M:dynamic((modeh/2,modeh/4,fixed_rule/3,banned/2,lookahead/2, 1207 lookahead_cons/2,lookahead_cons_var/2,'$prob'/2,output/1,input/1,input_cw/1, 1208 ref_clause/1,ref/1,model/1,neg/1,rule/5,determination/2, 1209 bg_on/0,bg/1,bgc/1,in_on/0,in/1,inc/1,int/1, 1210 query_rule/4, 1211 zero_clauses/1,tabled/1, 1212 fold/2)), 1213 retractall(M:tabled(_)), 1214 style_check(-discontiguous). 1215 1216user:term_expansion(end_of_file, end_of_file) :- 1217 lemur_file(Source), 1218 prolog_load_context(source, Source), 1219 retractall(lemur_file(Source)), 1220 prolog_load_context(module, M), 1221 lm_input_mod(M),!, 1222 make_dynamic(M), 1223 retractall(lm_input_mod(M)), 1224 style_check(+discontiguous). 1225 1226user:term_expansion(In, Out) :- 1227 \+ current_prolog_flag(xref, true), 1228 lemur_file(Source), 1229 prolog_load_context(source, Source), 1230 lemur_expansion(In, Out)