more CLP(BN) fixes, add softening option when computing MLE estimators.

CLPBN/clpbn.yap

@@ -78,7 +78,7 @@
 	       influences/4
 	      ]).
 
-:- dynamic solver/1,output/1,use/1,suppress_attribute_display/1.
+:- dynamic solver/1,output/1,use/1,suppress_attribute_display/1, parameter_softening/1.
 
 solver(jt).
 
@@ -86,6 +86,7 @@ solver(jt).
 %output(gviz(user_error)).
 output(no).
 suppress_attribute_display(false).
+parameter_softening(laplace).
 
 clpbn_flag(Flag,Option) :-
 	clpbn_flag(Flag, Option, Option).
@@ -111,6 +112,9 @@ clpbn_flag(bnt_model,Before,After) :-
 clpbn_flag(suppress_attribute_display,Before,After) :-
 	retract(suppress_attribute_display(Before)),
 	assert(suppress_attribute_display(After)).
+clpbn_flag(parameter_softening,Before,After) :-
+	retract(parameter_softening(Before)),
+	assert(parameter_softening(After)).
 
 
 {Var = Key with Dist} :-

CLPBN/clpbn/examples/School/tables.yap

@@ -37,29 +37,12 @@ rating_prob_table([0.9,0.05,0.01,
 		   0.09,0.9,0.09,
 		   0.01,0.05,0.9]).
 
-/*
-build_sats_table(LSats, Key, Table) :-
-	cpt_average(LSats, Key, [h,m,l], AggTable),
-	{ V = tmp1(Key) with AggTable},
-	rating_prob_table(SatsTable),
-	Table = p([h,m,l], SatsTable, [V]).
-*/
-
 build_rating_table(LSats, Key, Table) :-
 	cpt_average(LSats, Key, [h,m,l], 1.00, Table).
 
 build_grades_table(LGrades, Key, Table) :-
 	cpt_average(LGrades, Key, [a,b,c,d], 1.00, Table).
 
-/*
-build_grades_table(LGrades, Key, Table) :-
-	cpt_average(LGrades, Key, [a,b,c,d], AggTable),
-	{ V = tmp2(Key) with AggTable},
-	rating_prob_table(Tab),
-	Table = p([a,b,c,d], Tab, [V]).
-*/
-
-
 abi_table( _, [0.50, 0.40,  0.10]).
 
 

CLPBN/clpbn/gibbs.yap

@@ -32,7 +32,7 @@
 	project_from_CPT/3,
 	reorder_CPT/5,
 	multiply_possibly_deterministic_factors/3,
-	row_from_possibly_deterministic_CPT/3,
+	column_from_possibly_deterministic_CPT/3,
 	normalise_possibly_deterministic_CPT/2,
 	list_from_CPT/2]).
 
@@ -88,6 +88,7 @@ initialise(LVs, Graph, GVs, OutputVars, VarOrder) :-
 	graph_representation(LVs, Graph, 0, Keys, TGraph),
 	compile_graph(Graph),
 	topsort(TGraph, VarOrder),
+%writeln(TGraph:VarOrder),
 %	show_sorted(VarOrder, Graph),
 	add_all_output_vars(GVs, Keys, OutputVars).
 
@@ -108,7 +109,6 @@ graph_representation([V|Vs], Graph, I0, Keys, TGraph) :-
 	clpbn:get_atts(V,[evidence(_)]), !,
 	clpbn:get_atts(V, [dist(Id,Parents)]),
 	get_possibly_deterministic_dist_matrix(Id, Parents, _, Vals, Table),
-matrix:matrix_to_list(Table,T),writeln(T),
 	get_sizes(Parents, Szs),
 	length(Vals,Sz),
 	project_evidence_out([V|Parents],[V|Parents],Table,[Sz|Szs],Variables,NewTable),
@@ -238,7 +238,6 @@ mult_list([Sz|Sizes],Mult0,Mult) :-
 % compile node as set of facts, faster execution 
 compile_var(TotSize,I,_Vals,Sz,CPTs,Parents,_Sizes,Graph) :-
 	TotSize < 1024*64, TotSize > 0, !,
-writeln(I), (I=55->assert(a); retractall(a)),
 	multiply_all(I,Parents,CPTs,Sz,Graph).
 % do it dynamically
 compile_var(_,_,_,_,_,_,_,_).
@@ -247,7 +246,6 @@ multiply_all(I,Parents,CPTs,Sz,Graph) :-
 	markov_blanket_instance(Parents,Graph,Values),
 	(
 	    multiply_all(CPTs,Graph,Probs)
-, (a->writeln(Probs);true)
 	->
 	    store_mblanket(I,Values,Probs)
 	;
@@ -272,12 +270,9 @@ fetch_val([_|Vals],I0,Pos) :-
 	I is I0+1,
 	fetch_val(Vals,I,Pos).
 
-:- dynamic a/0.
-
 multiply_all([tabular(Table,_,Parents)|CPTs],Graph,Probs) :-
 	fetch_parents(Parents, Graph, Vals),
-	row_from_possibly_deterministic_CPT(Table,Vals,Probs0),
-	(a  -> list_from_CPT(Probs0,LProbs0), writeln(s:LProbs0) ; true),
+	column_from_possibly_deterministic_CPT(Table,Vals,Probs0),
 	multiply_more(CPTs,Graph,Probs0,Probs).
 
 fetch_parents([], _, []).
@@ -288,12 +283,10 @@ fetch_parents([P|Parents], Graph, [Val|Vals]) :-
 multiply_more([],_,Probs0,LProbs) :-
 	normalise_possibly_deterministic_CPT(Probs0, Probs),
 	list_from_CPT(Probs, LProbs0),
-	(a  -> writeln(e:LProbs0) ; true),
 	accumulate_up_list(LProbs0, 0.0, LProbs).
 multiply_more([tabular(Table,_,Parents)|CPTs],Graph,Probs0,Probs) :-
 	fetch_parents(Parents, Graph, Vals),
-	row_from_possibly_deterministic_CPT(Table, Vals, P0),
-	(a  -> list_from_CPT(P0, L0), list_from_CPT(Probs0, LI), writeln(m:LI:L0) ; true),
+	column_from_possibly_deterministic_CPT(Table, Vals, P0),
 	multiply_possibly_deterministic_factors(Probs0, P0, ProbsI),
 	multiply_more(CPTs,Graph,ProbsI,Probs).
 
@@ -379,7 +372,7 @@ gen_e0(Sz,[0|E0L]) :-
 
 process_chains(0,_,F,F,_,_,Est,Est) :- !.
 process_chains(ToDo,VarOrder,End,Start,Graph,Len,Est0,Estf) :-
-format('ToDo = ~d~n',[ToDo]),
+%format('ToDo = ~d~n',[ToDo]),
 	process_chains(Start,VarOrder,Int,Graph,Len,Est0,Esti),
 % (ToDo mod 100 =:= 1 -> statistics,cvt2problist(Esti, Probs), Int =[S|_], format('did ~d: ~w~n ~w~n',[ToDo,Probs,S]) ; true),
 	ToDo1 is ToDo-1,
@@ -407,11 +400,9 @@ do_var(I,Sample,Sample0,Graph) :-
 	;
 	  arg(I,Graph,var(_,_,_,_,_,CPTs,Parents,_,_)),
 	  fetch_parents(Parents,I,Sample,Sample0,Bindings),
-CPTs=[tabular(T,_,_)|_], matrix:matrix_dims(T,Dims), writeln(I:1:Bindings:Dims),
 	  multiply_all_in_context(Parents,Bindings,CPTs,Graph,Vals)
 	),
 	X is random,
-writeln(I:X:Vals),
 	pick_new_value(Vals,X,0,Val),
 	arg(I,Sample,Val).
 

CLPBN/clpbn/matrix_cpt_utils.yap

@@ -11,10 +11,10 @@
 	   unit_CPT/2,
 	   sum_out_from_CPT/4,
 	   list_from_CPT/2,
-	   row_from_CPT/3,
+	   column_from_CPT/3,
 	   multiply_factors/3,
 	   normalise_possibly_deterministic_CPT/2,
-	   row_from_possibly_deterministic_CPT/3,
+	   column_from_possibly_deterministic_CPT/3,
 	   multiply_possibly_deterministic_factors/3]).
 
 :- use_module(dists,
@@ -40,7 +40,7 @@
 	       matrix_to_list/2,
 	       matrix_agg_lines/3,
 	       matrix_op_to_lines/4,
-	       matrix_row/3]).
+	       matrix_column/3]).
 
 init_CPT(List, Sizes, TAB) :-
 	matrix_new(floats, Sizes, List, TAB),
@@ -212,12 +212,12 @@ zero_map([]).
 zero_map([0|Zeros]) :-
 	zero_map(Zeros).
 
-row_from_CPT(CPT, Parents, Row) :-
-	matrix_row(CPT, Parents, Row),
-	matrix_to_logs(Row).
+col_from_CPT(CPT, Parents, Column) :-
+	matrix_col(CPT, Parents, Column),
+	matrix_to_logs(Column).
 
-row_from_possibly_deterministic_CPT(CPT, Parents, Row) :-
-	matrix_row(CPT, Parents, Row).
+column_from_possibly_deterministic_CPT(CPT, Parents, Column) :-
+	matrix_column(CPT, Parents, Column).
 
 multiply_factors(F1, F2, F) :-
 	matrix_op(F1,F2,+,F).

CLPBN/learning/em.yap

@@ -23,7 +23,8 @@
 	      [run_all/1,
 	       clpbn_vars/2,
 	       normalise_counts/2,
-	       compute_likelihood/3]).
+	       compute_likelihood/3,
+	       soften_sample/2]).
 
 :- use_module(library(lists),
 	      [member/2]).
@@ -65,7 +66,7 @@ init_em(Items, state(AllVars, AllDists, AllDistInstances, MargVars)) :-
 em_loop(Its, Likelihood0, State, MaxError, MaxIts, LikelihoodF, FTables) :-
 	estimate(State, LPs),
 	maximise(State, Tables, LPs, Likelihood),
-	writeln(Likelihood:Tables),
+	writeln(Likelihood:Likelihood0:Tables),
 	(
 	    (
 	     (Likelihood - Likelihood0)/Likelihood < MaxError
@@ -152,7 +153,8 @@ compute_parameters([], [], [], Lik, Lik).
 compute_parameters([Id-Samples|Dists], [Id-NewTable|Tables], Ps, Lik0, Lik) :-
 	empty_dist(Id, Table0),
 	add_samples(Samples, Table0, Ps, MorePs),
-	normalise_counts(Table0, NewTable),
+	soften_sample(Table0, SoftenedTable),
+	normalise_counts(SoftenedTable, NewTable),
 	compute_likelihood(Table0, NewTable, DeltaLik),
 	dist_new_table(Id, NewTable),
 	NewLik is Lik0+DeltaLik,

CLPBN/learning/learn_utils.yap

@@ -5,15 +5,23 @@
 :- module(clpbn_learn_utils, [run_all/1,
 			      clpbn_vars/2,
 			      normalise_counts/2,
-			      compute_likelihood/3]).
+			      compute_likelihood/3,
+			      soften_sample/2,
+			      soften_sample/3]).
+
+:- use_module(library(clpbn),
+	      [clpbn_flag/2]).
 
 :- use_module(library(matrix),
 	      [matrix_agg_lines/3,
 	       matrix_op_to_lines/4,
+	       matrix_agg_cols/3,
+	       matrix_op_to_cols/4,
 	       matrix_to_logs/2,
 	       matrix_op/4,
 	       matrix_sum/2,
-	       matrix_to_list/2]).
+	       matrix_to_list/2,
+	       matrix_op_to_all/4]).
 
 :- meta_predicate run_all(:).
 
@@ -50,6 +58,25 @@ get_var_has_same_key([K-V|KVs],K,V,KVs0)  :- !,
 	get_var_has_same_key(KVs,K,V,KVs0).
 get_var_has_same_key(KVs,_,_,KVs).
 
+soften_sample(T0,T) :-
+	clpbn_flag(parameter_softening, Soften),
+	soften_sample(Soften, T0, T).
+
+soften_sample(no,T,T).
+soften_sample(m_estimate(M), T0, T) :-
+	matrix_agg_cols(T0,+,Cols),matrix:matrix_to_list(Cols), writeln(Cols),
+	matrix_op_to_all(Cols, *, M, R),
+	matrix_op_to_cols(T0,+,R,T).
+soften_sample(auto_m, T0,T) :-
+	matrix_agg_cols(T0,+,Cols),matrix:matrix_to_list(Cols), writeln(Cols),
+	matrix_sum(Cols,TotM),
+	M is sqrt(TotM),
+	matrix_op_to_all(Cols, *, M, R),
+	matrix_op_to_cols(T0,+,R,T).
+soften_sample(laplace,T0,T) :-
+	matrix_op_to_all(T0, +, 1, T).
+
+
 normalise_counts(MAT,NMAT) :-
 	matrix_agg_lines(MAT, +, Sum),
 	matrix_op_to_lines(MAT, Sum, /, NMAT).

CLPBN/learning/mle.yap

@@ -14,6 +14,8 @@
 :- use_module(library('clpbn/learning/learn_utils'),
 	      [run_all/1,
 	       clpbn_vars/2,
+	       normalise_counts/2,
+	       soften_table/2,
 	       normalise_counts/2]).
 
 :- use_module(library('clpbn/dists'),
@@ -21,8 +23,7 @@
 	       dist_new_table/2]).
 
 :- use_module(library(matrix),
-	      [matrix_inc/2,
-	       matrix_op_to_all/4]).
+	      [matrix_inc/2]).
 
 
 learn_parameters(Items, Tables) :-
@@ -89,12 +90,14 @@ compute_tables(Parameters, Sample, NewTables) :-
 	add_priors(Parameters, Tables, NewTables).
 
 estimator([], []).
-estimator([val(Id,Sample)|Samples], [NewTable|Tables]) :-
+estimator([val(Id,Sample)|Samples], [NewDist|Tables]) :-
 	empty_dist(Id, NewTable),
 	id_samples(Id, Samples, IdSamples, MoreSamples),
 	mle([Sample|IdSamples], NewTable),
+	soften_table(NewTable, SoftenedTable),
+	normalise_counts(SoftenedTable, NewDist),
 	% replace matrix in distribution
-	dist_new_table(Id, NewTable),
+	dist_new_table(Id, NewDist),
 	estimator(MoreSamples, Tables).
 
 
@@ -108,25 +111,3 @@ mle([Sample|IdSamples], Table) :-
 	mle(IdSamples, Table).
 mle([], _).
 
-add_priors([], Tables, NewTables) :-
-	normalise(Tables, NewTables).
-add_priors([laplace|_], Tables, NewTables) :- !,
-	laplace(Tables, TablesI),
-	normalise(TablesI, NewTables).
-add_priors([m_estimate(M)|_], Tables, NewTables) :- !,
-	add_mestimate(Tables, M, TablesI),
-	normalise(TablesI, NewTables).
-add_priors([_|Parms], Tables, NewTables) :-
-	add_priors(Parms, Tables, NewTables).
-
-normalise([], []).
-normalise([T0|TablesI], [T|NewTables]) :-
-	normalise_counts(T0, T),
-	normalise(TablesI, NewTables).
-
-laplace([], []).
-laplace([T0|TablesI], [T|NewTables]) :-
-	matrix_op_to_all(T0, +, 1, T),
-	laplace(TablesI, NewTables).
-
-

library/matrix.yap

@@ -82,7 +82,7 @@ typedef enum {
 	    matrix_set_all_that_disagree/5,
 	    matrix_expand/3,
 	    matrix_select/4,
-	    matrix_row/3
+	    matrix_column/3
 	    ]).
 
 :- load_foreign_files([matrix], [], init_matrix).

library/matrix/matrix.c

@@ -2299,7 +2299,7 @@ matrix_select(void)
 /* given a matrix M and a set of N-1 dims, get the first dimension
 */ 
 static int
-matrix_row(void)
+matrix_column(void)
 {
   int size, i, ndims, newdims[1];
   int indx[MAX_DIMS];
@@ -2956,7 +2956,7 @@ init_matrix(void)
   YAP_UserCPredicate("matrix_shuffle", matrix_transpose, 3);
   YAP_UserCPredicate("matrix_expand", matrix_expand, 3);
   YAP_UserCPredicate("matrix_select", matrix_select, 4);
-  YAP_UserCPredicate("matrix_row", matrix_row, 3);
+  YAP_UserCPredicate("matrix_column", matrix_column, 3);
   YAP_UserCPredicate("matrix_to_logs", matrix_log_all,1);
   YAP_UserCPredicate("matrix_to_exps", matrix_exp_all, 1);
   YAP_UserCPredicate("matrix_to_logs", matrix_log_all2,2);