learning is debugging

packages/CLPBN/clpbn.yap

@@ -122,7 +122,7 @@
 :- dynamic solver/1,output/1,use/1,suppress_attribute_display/1, parameter_softening/1, em_solver/1, use_parfactors/1.
 
 solver(ve).
-em_solver(ve).
+em_solver(bp).
 
 :- meta_predicate probability(:,-), conditional_probability(:,:,-).
 

packages/CLPBN/clpbn/ground_factors.yap

@@ -61,7 +61,7 @@ ground_all_keys([], _).
 ground_all_keys([V|GVars], AllKeys) :-
 	clpbn:get_atts(V,[key(Key)]), 
 	\+ ground(Key), !,
-wroteln(g:Key),
+writeln(g:Key),
 	member(Key, AllKeys),
 	ground_all_keys(GVars, AllKeys).
 ground_all_keys([_V|GVars], AllKeys) :-

packages/CLPBN/clpbn/horus_ground.yap

@@ -52,25 +52,40 @@
 
 
 call_horus_ground_solver(QueryVars, QueryKeys, AllKeys, Factors, Evidence, Output) :-
-  b_hash_new(Hash0),
-  keys_to_ids(AllKeys, 0, Hash0, Hash),
-  get_factors_type(Factors, Type),
-  evidence_to_ids(Evidence, Hash, EvidenceIds),
-  factors_to_ids(Factors, Hash, FactorIds),
-  %writeln(type:Type), writeln(''),
-  %writeln(allKeys:AllKeys), writeln(''),
-  %sort(AllKeys,SKeys),writeln(allKeys:SKeys), writeln(''),
-  %writeln(factors:Factors), writeln(''),
-  %writeln(factorIds:FactorIds), writeln(''),
-  %writeln(evidence:Evidence), writeln(''),
-  %writeln(evidenceIds:EvidenceIds), writeln(''),
-  cpp_create_ground_network(Type, FactorIds, EvidenceIds, Network),
-  %get_vars_information(AllKeys, StatesNames),
-  %terms_to_atoms(AllKeys, KeysAtoms),
-  %cpp_set_vars_information(KeysAtoms, StatesNames),
-  run_solver(ground(Network,Hash), QueryKeys, Solutions),
-  clpbn_bind_vals([QueryVars], Solutions, Output),
-  cpp_free_ground_network(Network).
+	call_horus_ground_solver_for_probabilities(QueryKeys, AllKeys, Factors, Evidence, Solutions),
+	clpbn_bind_vals([QueryVars], Solutions, Output).
+
+call_horus_ground_solver_for_probabilities(QueryKeys, _AllKeys, Factors, Evidence, Solutions) :-
+	attributes:all_attvars(AVars),
+	keys(AVars, AllKeys),
+	b_hash_new(Hash0),
+	keys_to_ids(AllKeys, 0, Hash0, Hash),
+	get_factors_type(Factors, Type),
+	evidence_to_ids(Evidence, Hash, EvidenceIds),
+	factors_to_ids(Factors, Hash, FactorIds),
+	writeln(queryKeys:QueryKeys), writeln(''),
+	writeln(type:Type), writeln(''),
+	writeln(allKeys:AllKeys), writeln(''),
+	sort(AllKeys,SKeys),writeln(allSortedKeys:SKeys), writeln(''),
+	keys_to_ids(SKeys, 0, Hash0, Hash),
+	writeln(factors:Factors), writeln(''),
+	writeln(factorIds:FactorIds), writeln(''),
+	writeln(evidence:Evidence), writeln(''),
+	writeln(evidenceIds:EvidenceIds), writeln(''),
+	cpp_create_ground_network(Type, FactorIds, EvidenceIds, Network),
+	get_vars_information(AllKeys, StatesNames),
+	terms_to_atoms(AllKeys, KeysAtoms),
+	cpp_set_vars_information(KeysAtoms, StatesNames),
+	run_solver(ground(Network,Hash), QueryKeys, Solutions),
+	cpp_free_ground_network(Network).
+
+
+keys([], []).
+keys([V|AVars], [K|AllKeys]) :-
+	clpbn:get_atts(V,[key(K)]), !,
+	keys(AVars, AllKeys).
+keys([_V|AVars], AllKeys) :-
+	keys(AVars, AllKeys).
 
 
 run_solver(ground(Network,Hash), QueryKeys, Solutions) :-
@@ -94,9 +109,13 @@ get_factors_type([f(markov, _, _, _)|_], markov) :- ! .
 
 
 list_of_keys_to_ids([], _, []).
+list_of_keys_to_ids([List|Extra], Hash, [IdList|More]) :-
+	List = [_|_], !,
+	list_of_keys_to_ids(List, Hash, IdList),
+	list_of_keys_to_ids(Extra, Hash, More).
 list_of_keys_to_ids([Key|QueryKeys], Hash, [Id|QueryIds]) :-
-  b_hash_lookup(Key, Id, Hash),
-  list_of_keys_to_ids(QueryKeys, Hash, QueryIds).
+	b_hash_lookup(Key, Id, Hash),
+	list_of_keys_to_ids(QueryKeys, Hash, QueryIds).
 
 
 factors_to_ids([], _, []).
@@ -134,10 +153,33 @@ terms_to_atoms(K.Ks, Atom.As) :-
 finalize_horus_ground_solver(bp(Network, _)) :-
   cpp_free_ground_network(Network).
 
+%
+% QVars: all query variables?
+% 
+% 
+init_horus_ground_solver(QueryVars, _AllVars, _, horus(GKeys, Keys, Factors, Evidence)) :-
+	trace,
+	generate_networks(QueryVars, GKeys, [], Keys, [], Factors, [], Evidence),
+	writeln(qvs:QueryVars),
+	writeln(Keys), !.
 
-init_horus_ground_solver(_, _AllVars0, _, bp(_BayesNet, _DistIds)) :- !.
+%
+% as you add query vars the network grows
+% until you reach the last variable.
+%
+generate_networks([QVars|QueryVars], [GK|GKeys], _K0, K, _F0, F, _E0, E) :-
+	clpbn:generate_network(QVars, GK, KI, FI, EI),
+	generate_networks(QueryVars, GKeys, KI, K, FI, F, EI, E).
+generate_networks([], [], K, K, F, F, E, E).
 
-run_horus_ground_solver(_QueryVars, _Solutions, bp(_Network, _DistIds)) :- !.
+%
+% just call horus solver.
+%
+run_horus_ground_solver(_QueryVars, Solutions, horus(GKeys, Keys, Factors, Evidence) ) :- !,
+	writeln(sols:Solutions),
+	writeln(state:_State),
+	trace,
+	call_horus_ground_solver_for_probabilities(GKeys, Keys, Factors, Evidence, Solutions).
 
 %bp([[]],_,_) :- !.
 %bp([QueryVars], AllVars, Output) :-

packages/CLPBN/clpbn/ve.yap

@@ -81,7 +81,6 @@ ve([LVs],Vs0,AllDiffs) :-
 init_ve_solver(Qs, Vs0, _, LVis) :-
 	check_for_agg_vars(Vs0, Vs1),
 	% LVi will have a  list of CLPBN variables
-	% Tables0 will have the full data on each variable
 	init_influences(Vs1, G, RG),
 	init_ve_solver_for_questions(Qs, G, RG, _, LVis).
 

packages/CLPBN/examples/learning/prof_params.pfl

@@ -0,0 +1,56 @@
+% learn distribution for school database.
+
+:- use_module(library(pfl)).
+
+:- use_module(library(clpbn/learning/em)).
+
+bayes abi(K)::[h,m,l] ; abi_table ; [professor(K)].
+
+bayes pop(K)::[h,m,l], abi(K) ; pop_table ; [professor(K)].
+
+abi_table([0.3,0.3,0.4]).
+
+pop_table([0.3,0.3,0.4,0.3,0.3,0.4,0.3,0.3,0.4]).
+
+goal_list([/*abi(p0,h),
+	abi(p1,m),
+	abi(p2,m),
+	abi(p3,m),
+	abi(p4,l),*/
+	pop(p5,h),
+	abi(p5,_),
+	abi(p6,_),
+	pop(p7,_)]).
+
+professor(p1).
+professor(p2).
+professor(p3).
+professor(p4).
+professor(p5).
+professor(p6).
+professor(p7).
+professor(p8).
+
+%:- clpbn:set_clpbn_flag(em_solver,gibbs).
+%:- clpbn:set_clpbn_flag(em_solver,jt).
+:- clpbn:set_clpbn_flag(em_solver,ve).
+%:- clpbn:set_clpbn_flag(em_solver,bp).
+
+timed_main :-
+	statistics(runtime, _),
+	main(Lik),
+	statistics(runtime, [T,_]),
+	format('Took ~d msec and Lik ~3f~n',[T,Lik]).
+
+main(Lik) :-
+	goal_list(L),
+%	run_queries(L),
+        em(L,0.01,10,_,Lik).
+
+run_queries([]).
+run_queries(Q.L) :-
+	call(Q),
+	run_queries(L).
+
+
+

packages/CLPBN/learning/em.yap

@@ -97,7 +97,7 @@ init_em(Items, state( AllDists, AllDistInstances, MargVars, SolverVars)) :-
 em_loop(Its, Likelihood0, State, MaxError, MaxIts, LikelihoodF, FTables) :-
 	estimate(State, LPs),
 	maximise(State, Tables, LPs, Likelihood),
-%	writeln(Likelihood:Its:Likelihood0:Tables),
+	writeln(iteration:Its:Likelihood:Its:Likelihood0:Tables),
 	(
 	    (
 	     abs((Likelihood - Likelihood0)/Likelihood) < MaxError
@@ -166,6 +166,14 @@ find_variables([K|PKeys], AllVars0, [Parent|Parents]) :-
 	find_variable(K, AllVars0, Parent),
 	find_variables(PKeys, AllVars0, Parents).
 
+%
+% in clp(bn) the whole network is constructed when you evaluate EM. In
+% pfl, we want to delay execution until as late as possible.
+% we just create a new variable and hope for the best.
+%
+%
+find_variable(K, [], Parent) :-
+	clpbn:put_atts(Parent, [key(K)]).
 find_variable(K, [Parent|_AllVars0], Parent) :-
 	clpbn:get_atts(Parent, [key(K0)]), K0 =@= K, !.
 find_variable(K, [_|AllVars0], Parent) :-