This repository has been archived on 2023-08-20. You can view files and clone it, but cannot push or open issues or pull requests.
yap-6.3/packages/CLPBN/clpbn.yap

630 lines
17 KiB
Prolog

:- module(clpbn, [{}/1,
clpbn_flag/2,
set_clpbn_flag/2,
clpbn_flag/3,
clpbn_key/2,
clpbn_init_solver/4,
clpbn_run_solver/3,
clpbn_finalize_solver/1,
clpbn_init_solver/5,
clpbn_run_solver/4,
clpbn_init_graph/1,
probability/2,
conditional_probability/3,
use_parfactors/1,
op( 500, xfy, with)]).
:- use_module(library(atts)).
:- use_module(library(bhash)).
:- use_module(library(lists)).
:- use_module(library(terms)).
%
% avoid the overhead of using goal_expansion/2.
%
:- multifile
user:term_expansion/2.
:- dynamic
user:term_expansion/2.
:- attribute key/1, dist/2, evidence/1.
:- use_module('clpbn/ve',
[ve/3,
check_if_ve_done/1,
init_ve_solver/4,
run_ve_solver/3
]).
:- use_module('clpbn/horus_ground',
[call_horus_ground_solver/6,
check_if_horus_ground_solver_done/1,
init_horus_ground_solver/4,
run_horus_ground_solver/3,
finalize_horus_ground_solver/1
]).
:- use_module('clpbn/horus_lifted',
[call_horus_lifted_solver/3,
check_if_horus_lifted_solver_done/1,
init_horus_lifted_solver/4,
run_horus_lifted_solver/3,
finalize_horus_lifted_solver/1
]).
:- use_module('clpbn/jt',
[jt/3,
init_jt_solver/4,
run_jt_solver/3
]).
:- use_module('clpbn/bdd',
[bdd/3,
init_bdd_solver/4,
run_bdd_solver/3
]).
%% :- use_module('clpbn/bnt',
%% [do_bnt/3,
%% check_if_bnt_done/1
%% ]).
:- use_module('clpbn/gibbs',
[gibbs/3,
check_if_gibbs_done/1,
init_gibbs_solver/4,
run_gibbs_solver/3
]).
:- use_module('clpbn/pgrammar',
[init_pcg_solver/4,
run_pcg_solver/3,
pcg_init_graph/0
]).
:- use_module('clpbn/graphs',
[
clpbn2graph/1
]).
:- use_module('clpbn/dists',
[
dist/4,
get_dist/4,
get_evidence_position/3,
get_evidence_from_position/3,
additive_dists/6
]).
:- use_module('clpbn/evidence',
[
store_evidence/1,
add_stored_evidence/2,
incorporate_evidence/2,
check_stored_evidence/2,
put_evidence/2
]).
:- use_module('clpbn/utils',
[
sort_vars_by_key/3
]).
:- use_module('clpbn/graphviz',
[clpbn2gviz/4]).
:- use_module(clpbn/ground_factors,
[generate_network/5]).
:- dynamic solver/1,output/1,use/1,suppress_attribute_display/1, parameter_softening/1, em_solver/1, use_parfactors/1.
solver(ve).
em_solver(bp).
:- meta_predicate probability(:,-), conditional_probability(:,:,-).
%output(xbif(user_error)).
%output(gviz(user_error)).
output(no).
suppress_attribute_display(false).
parameter_softening(m_estimate(10)).
use_parfactors(off).
clpbn_flag(Flag,Option) :-
clpbn_flag(Flag, Option, Option).
set_clpbn_flag(Flag,Option) :-
clpbn_flag(Flag, _, Option).
clpbn_flag(output,Before,After) :-
retract(output(Before)),
assert(output(After)).
clpbn_flag(solver,Before,After) :-
retract(solver(Before)),
assert(solver(After)).
clpbn_flag(em_solver,Before,After) :-
retract(em_solver(Before)),
assert(em_solver(After)).
clpbn_flag(bnt_solver,Before,After) :-
retract(bnt:bnt_solver(Before)),
assert(bnt:bnt_solver(After)).
clpbn_flag(bnt_path,Before,After) :-
retract(bnt:bnt_path(Before)),
assert(bnt:bnt_path(After)).
clpbn_flag(bnt_model,Before,After) :-
retract(bnt:bnt_model(Before)),
assert(bnt:bnt_model(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)).
clpbn_flag(use_factors,Before,After) :-
retract(use_parfactors(Before)),
assert(use_parfactors(After)).
{_} :-
solver(none), !.
{Var = Key with Dist} :-
put_atts(El,[key(Key),dist(DistInfo,Parents)]),
dist(Dist, DistInfo, Key, Parents),
add_evidence(Var,Key,DistInfo,El)
% ,writeln({Var = Key with Dist})
.
%
% make sure a query variable is reachable by the garbage collector.
%
% we use a mutable variable to avoid unnecessary trailing.
%
store_var(El) :-
nb_current(clpbn_qvars, Mutable),
nonvar(Mutable), !,
get_mutable(Tail, Mutable),
update_mutable(El.Tail, Mutable).
store_var(El) :-
init_clpbn_vars(El).
init_clpbn_vars(El) :-
create_mutable(El, Mutable),
b_setval(clpbn_qvars, Mutable).
check_constraint(Constraint, _, _, Constraint) :- var(Constraint), !.
check_constraint((A->D), _, _, (A->D)) :- var(A), !.
check_constraint((([A|B].L)->D), Vars, NVars, (([A|B].NL)->D)) :- !,
check_cpt_input_vars(L, Vars, NVars, NL).
check_constraint(Dist, _, _, Dist).
check_cpt_input_vars([], _, _, []).
check_cpt_input_vars([V|L], Vars, NVars, [NV|NL]) :-
replace_var(Vars, V, NVars, NV),
check_cpt_input_vars(L, Vars, NVars, NL).
replace_var([], V, [], V).
replace_var([V|_], V0, [NV|_], NV) :- V == V0, !.
replace_var([_|Vars], V, [_|NVars], NV) :-
replace_var(Vars, V, NVars, NV).
add_evidence(V,Key,Distinfo,NV) :-
nonvar(V), !,
get_evidence_position(V, Distinfo, Pos),
check_stored_evidence(Key, Pos),
store_var(NV),
clpbn:put_atts(NV,evidence(Pos)).
add_evidence(V,K,_,V) :-
add_stored_evidence(K,V),
store_var(V).
clpbn_marginalise(V, Dist) :-
attributes:all_attvars(AVars),
project_attributes([V], AVars),
clpbn_display:get_atts(V, posterior(_,_,Dist,_)).
%
% called by top-level
% or by call_residue/2
%
project_attributes(GVars, _AVars0) :-
use_parfactors(on),
clpbn_flag(solver, Solver), Solver \= fove, !,
generate_network(GVars, GKeys, Keys, Factors, Evidence),
call_ground_solver(Solver, GVars, GKeys, Keys, Factors, Evidence).
project_attributes(GVars, AVars) :-
suppress_attribute_display(false),
AVars = [_|_],
solver(Solver),
( GVars = [_|_] ; Solver = graphs), !,
clpbn_vars(AVars, DiffVars, AllVars),
get_clpbn_vars(GVars,CLPBNGVars0),
simplify_query_vars(CLPBNGVars0, CLPBNGVars),
(output(xbif(XBifStream)) -> clpbn2xbif(XBifStream,ve,AllVars) ; true),
(output(gviz(XBifStream)) -> clpbn2gviz(XBifStream,sort,AllVars,GVars) ; true),
(
Solver = graphs
->
write_out(Solver, [[]], AllVars, DiffVars)
;
write_out(Solver, [CLPBNGVars], AllVars, DiffVars)
).
project_attributes(_, _).
match([], _Keys).
match([V|GVars], Keys) :-
clpbn:get_atts(V,[key(GKey)]),
member(GKey, Keys), ground(GKey), !,
match(GVars, Keys).
match([_V|GVars], Keys) :-
match(GVars, Keys).
clpbn_vars(AVars, DiffVars, AllVars) :-
sort_vars_by_key(AVars,SortedAVars,DiffVars),
incorporate_evidence(SortedAVars, AllVars).
get_clpbn_vars([V|GVars],[V|CLPBNGVars]) :-
get_atts(V, [key(_)]), !,
get_clpbn_vars(GVars,CLPBNGVars).
get_clpbn_vars([_|GVars],CLPBNGVars) :-
get_clpbn_vars(GVars,CLPBNGVars).
get_clpbn_vars([],[]).
get_clpbn_vars([V|GVars],[V|CLPBNGVars]) :-
get_atts(V, [key(_)]), !,
get_clpbn_vars(GVars,CLPBNGVars).
get_clpbn_vars([_|GVars],CLPBNGVars) :-
get_clpbn_vars(GVars,CLPBNGVars).
simplify_query_vars(LVs0, LVs) :-
sort(LVs0,LVs1),
get_rid_of_ev_vars(LVs1,LVs).
%
% some variables might already have evidence in the data-base.
%
get_rid_of_ev_vars([],[]).
get_rid_of_ev_vars([V|LVs0],LVs) :-
clpbn:get_atts(V, [dist(Id,_),evidence(Pos)]), !,
get_evidence_from_position(Ev, Id, Pos),
clpbn_display:put_atts(V, [posterior([],Ev,[],[])]), !,
get_rid_of_ev_vars(LVs0,LVs).
get_rid_of_ev_vars([V|LVs0],[V|LVs]) :-
get_rid_of_ev_vars(LVs0,LVs).
% do nothing if we don't have query variables to compute.
write_out(graphs, _, AVars, _) :-
clpbn2graph(AVars).
write_out(ve, GVars, AVars, DiffVars) :-
ve(GVars, AVars, DiffVars).
write_out(jt, GVars, AVars, DiffVars) :-
jt(GVars, AVars, DiffVars).
write_out(bdd, GVars, AVars, DiffVars) :-
bdd(GVars, AVars, DiffVars).
write_out(bp, _GVars, _AVars, _DiffVars) :-
writeln('interface not supported anymore').
%bp(GVars, AVars, DiffVars).
write_out(gibbs, GVars, AVars, DiffVars) :-
gibbs(GVars, AVars, DiffVars).
write_out(bnt, GVars, AVars, DiffVars) :-
do_bnt(GVars, AVars, DiffVars).
write_out(fove, GVars, AVars, DiffVars) :-
call_horus_lifted_solver(GVars, AVars, DiffVars).
% call a solver with keys, not actual variables
call_ground_solver(bp, GVars, GoalKeys, Keys, Factors, Evidence) :- !,
call_horus_ground_solver(GVars, GoalKeys, Keys, Factors, Evidence, _Answ).
call_ground_solver(Solver, GVars, _GoalKeys, Keys, Factors, Evidence) :-
% traditional solver
b_hash_new(Hash0),
gvars_in_hash(GVars,Hash0, HashI),
keys_to_vars(Keys, AllVars, HashI, Hash1),
evidence_to_vars(Evidence, _EVars, Hash1, Hash),
factors_to_dists(Factors, Hash),
% evidence
retract(use_parfactors(on)),
write_out(Solver, [GVars], AllVars, _),
assert(use_parfactors(on)).
%
% convert a PFL network (without constriants)
% into CLP(BN) for evaluation
%
gvars_in_hash([V|GVars],Hash0, Hash) :-
get_atts(V, [key(K)]),
b_hash_insert(Hash0, K, V, HashI),
gvars_in_hash(GVars,HashI, Hash).
gvars_in_hash([],Hash, Hash).
keys_to_vars([], [], Hash, Hash).
keys_to_vars([K|Keys], [V|Vs], Hash0, Hash) :-
b_hash_lookup(K, V, Hash0), !,
keys_to_vars(Keys, Vs, Hash0, Hash).
keys_to_vars([K|Keys], [V|Vs],Hash0, Hash) :-
b_hash_insert(Hash0, K, V, HashI),
keys_to_vars(Keys, Vs, HashI, Hash).
evidence_to_vars([], [], Hash, Hash).
evidence_to_vars([K=E|Keys], [V|Vs], Hash0, Hash) :-
b_hash_lookup(K, V, Hash0), !,
clpbn:put_atts(V,[evidence(E)]),
evidence_to_vars(Keys, Vs, Hash0, Hash).
evidence_to_vars([K=E|Keys], [V|Vs],Hash0, Hash) :-
b_hash_insert(Hash0, K, V, HashI),
clpbn:put_atts(V,[evidence(E)]),
evidence_to_vars(Keys, Vs, HashI, Hash).
factors_to_dists([], _Hash).
factors_to_dists([f(bayes,_Id,Ks,CPT)|Factors], Hash) :-
keys_to_vars(Ks, Hash, [V|Parents]),
Ks =[Key|_],
pfl:skolem(Key, Domain),
dist(p(Domain,CPT,Parents), DistInfo, Key, Parents),
put_atts(V,[dist(DistInfo,Parents)]),
factors_to_dists(Factors, Hash).
keys_to_vars([], _Hash, []).
keys_to_vars([K|Ks], Hash, [V|Vs]) :-
b_hash_lookup(K,V,Hash),
keys_to_vars(Ks, Hash, Vs).
get_bnode(Var, Goal) :-
get_atts(Var, [key(Key),dist(Dist,Parents)]),
get_dist(Dist,_,Domain,CPT),
(Parents = [] -> X = tab(Domain,CPT) ; X = tab(Domain,CPT,Parents)),
dist_goal(X, Key, Goal0),
include_evidence(Var, Goal0, Key, Goal).
include_evidence(Var, Goal0, Key, ((Key:-Ev),Goal0)) :-
get_atts(Var, [evidence(Ev)]), !.
include_evidence(_, Goal0, _, Goal0).
dist_goal(Dist, Key, (Key=NDist)) :-
term_variables(Dist, DVars),
process_vars(DVars, DKeys),
my_copy_term(Dist,DVars, NDist,DKeys).
my_copy_term(V, DVars, Key, DKeys) :-
find_var(DVars, V, Key, DKeys).
my_copy_term(A, _, A, _) :- atomic(A), !.
my_copy_term(T, Vs, NT, Ks) :-
T =.. [Na|As],
my_copy_terms(As, Vs, NAs, Ks),
NT =.. [Na|NAs].
my_copy_terms([], _, [], _).
my_copy_terms([A|As], Vs, [NA|NAs], Ks) :-
my_copy_term(A, Vs, NA, Ks),
my_copy_terms(As, Vs, NAs, Ks).
find_var([V1|_], V, Key, [Key|_]) :- V1 == V, !.
find_var([_|DVars], V, Key, [_|DKeys]) :-
find_var(DVars, V, Key, DKeys).
process_vars([], []).
process_vars([V|Vs], [K|Ks]) :-
process_var(V, K),
process_vars(Vs, Ks).
process_var(V, K) :- get_atts(V, [key(K)]), !.
% oops: this variable has no attributes.
process_var(V, _) :- throw(error(instantiation_error,clpbn(attribute_goal(V)))).
%
% unify a CLPBN variable with something.
%
verify_attributes(Var, T, Goal) :-
get_atts(Var, [key(Key),dist(Dist,Parents)]), !,
/* oops, someone trying to bind a clpbn constrained variable */
bind_clpbn(T, Var, Key, Dist, Parents, Goal).
verify_attributes(_, _, []).
bind_clpbn(T, Var, _, _, _, do_not_bind_variable([put_evidence(T,Var)])) :-
nonvar(T),
!.
bind_clpbn(T, Var, Key, Dist, Parents, []) :- var(T),
get_atts(T, [key(Key1),dist(Dist1,Parents1)]),
(
bind_clpbns(Key, Dist, Parents, Key1, Dist1, Parents1)
->
(
get_atts(T, [evidence(Ev1)]) ->
bind_evidence_from_extra_var(Ev1,Var)
;
get_atts(Var, [evidence(Ev)]) ->
bind_evidence_from_extra_var(Ev,T)
;
true
)
;
fail
).
bind_clpbn(_, Var, _, _, _, _, []) :-
use(bnt),
check_if_bnt_done(Var), !.
bind_clpbn(_, Var, _, _, _, _, []) :-
use(ve),
check_if_ve_done(Var), !.
bind_clpbn(_, Var, _, _, _, _, []) :-
use(bp),
check_if_horus_ground_solver_done(Var), !.
bind_clpbn(_, Var, _, _, _, _, []) :-
use(jt),
check_if_ve_done(Var), !.
bind_clpbn(_, Var, _, _, _, _, []) :-
use(bdd),
check_if_bdd_done(Var), !.
bind_clpbn(T, Var, Key0, _, _, _, []) :-
get_atts(Var, [key(Key)]), !,
(
Key = Key0 -> true
;
% let us not loose whatever we had.
put_evidence(T,Var)
).
fresh_attvar(Var, NVar) :-
get_atts(Var, LAtts),
put_atts(NVar, LAtts).
% I will now allow two CLPBN variables to be bound together.
%bind_clpbns(Key, Dist, Parents, Key, Dist, Parents).
bind_clpbns(Key, Dist, Parents, Key1, Dist1, Parents1) :-
Key == Key1, !,
get_dist(Dist,_Type,_Domain,_Table),
get_dist(Dist1,_Type1,_Domain1,_Table1),
Dist = Dist1,
Parents = Parents1.
bind_clpbns(Key, _, _, _, Key1, _, _, _) :-
Key\=Key1, !, fail.
bind_clpbns(_, _, _, _, _, _, _, _) :-
format(user_error, 'unification of two bayesian vars not supported~n', []).
same_parents([],[]).
same_parents([P|Parents],[P1|Parents1]) :-
same_node(P,P1),
same_parents(Parents,Parents1).
same_node(P,P1) :- P == P1, !.
same_node(P,P1) :-
get_atts( P,[key(K)]),
get_atts(P1,[key(K)]),
P = P1.
bind_evidence_from_extra_var(Ev1,Var) :-
get_atts(Var, [evidence(Ev0)]), !,
Ev0 = Ev1.
bind_evidence_from_extra_var(Ev1,Var) :-
put_atts(Var, [evidence(Ev1)]).
user:term_expansion((A :- {}), ( :- true )) :- !, % evidence
prolog_load_context(module, M),
store_evidence(M:A).
clpbn_key(Var,Key) :-
get_atts(Var, [key(Key)]).
%
% This is a routine to start a solver, called by the learning procedures (ie, em).
% LVs is a list of lists of variables one is interested in eventually marginalising out
% Vs0 gives the original graph
% AllDiffs gives variables that are not fully constrainted, ie, we don't fully know
% the key. In this case, we assume different instances will be bound to different
% values at the end of the day.
%
clpbn_init_solver(LVs, Vs0, VarsWithUnboundKeys, State) :-
solver(Solver),
clpbn_init_solver(Solver, LVs, Vs0, VarsWithUnboundKeys, State).
clpbn_init_solver(gibbs, LVs, Vs0, VarsWithUnboundKeys, State) :-
init_gibbs_solver(LVs, Vs0, VarsWithUnboundKeys, State).
clpbn_init_solver(ve, LVs, Vs0, VarsWithUnboundKeys, State) :-
init_ve_solver(LVs, Vs0, VarsWithUnboundKeys, State).
clpbn_init_solver(bp, LVs, Vs0, VarsWithUnboundKeys, State) :-
init_horus_ground_solver(LVs, Vs0, VarsWithUnboundKeys, State).
clpbn_init_solver(jt, LVs, Vs0, VarsWithUnboundKeys, State) :-
init_jt_solver(LVs, Vs0, VarsWithUnboundKeys, State).
clpbn_init_solver(bdd, LVs, Vs0, VarsWithUnboundKeys, State) :-
init_bdd_solver(LVs, Vs0, VarsWithUnboundKeys, State).
clpbn_init_solver(pcg, LVs, Vs0, VarsWithUnboundKeys, State) :-
init_pcg_solver(LVs, Vs0, VarsWithUnboundKeys, State).
%
% LVs is the list of lists of variables to marginalise
% Vs is the full graph
% Ps are the probabilities on LVs.
%
%
clpbn_run_solver(LVs, LPs, State) :-
solver(Solver),
clpbn_run_solver(Solver, LVs, LPs, State).
clpbn_run_solver(gibbs, LVs, LPs, State) :-
run_gibbs_solver(LVs, LPs, State).
clpbn_run_solver(ve, LVs, LPs, State) :-
run_ve_solver(LVs, LPs, State).
clpbn_run_solver(bp, LVs, LPs, State) :-
run_horus_ground_solver(LVs, LPs, State).
clpbn_run_solver(jt, LVs, LPs, State) :-
run_jt_solver(LVs, LPs, State).
clpbn_run_solver(bdd, LVs, LPs, State) :-
run_bdd_solver(LVs, LPs, State).
clpbn_run_solver(pcg, LVs, LPs, State) :-
run_pcg_solver(LVs, LPs, State).
add_keys(Key1+V1,_Key2,Key1+V1).
%
% only useful for probabilistic context free grammars
%
clpbn_init_graph(pcg) :- !,
pcg_init_graph.
clpbn_init_graph(_).
clpbn_finalize_solver(State) :-
solver(bp), !,
functor(State, _, Last),
arg(Last, State, Info),
finalize_horus_ground_solver(Info).
clpbn_finalize_solver(_State).
probability(Goal, Prob) :-
findall(Prob, do_probability(Goal, [], Prob), [Prob]).
conditional_probability(Goal, ListOfGoals, Prob) :-
\+ ground(ListOfGoals), !,
throw(error(ground(ListOfGoals),conditional_probability(Goal, ListOfGoals, Prob))).
conditional_probability(Goal, ListOfGoals, Prob) :-
findall(Prob, do_probability(Goal, ListOfGoals, Prob), [Prob]).
do_probability(Goal, ListOfGoals, Prob) :-
evidence_to_var(Goal, C, NGoal, V),
call_residue(run( ListOfGoals, NGoal), Vars), !,
match_probability(Vars, C, V, Prob).
run(ListOfGoals,Goal) :-
do(ListOfGoals),
call(Goal).
do(M:ListOfGoals) :-
do(ListOfGoals, M).
do([]).
do([], _M).
do(G.ListOfGoals, M) :-
M:G,
do(ListOfGoals, M).
evidence_to_var(M:Goal, C, M:VItem, V) :- !,
evidence_to_var(Goal, C, VItem, V).
evidence_to_var(Goal, C, VItem, V) :-
Goal =.. [L|Args],
variabilise_last(Args, C, NArgs, V),
VItem =.. [L|NArgs].
variabilise_last([Arg], Arg, [V], V).
variabilise_last([Arg1,Arg2|Args], Arg, Arg1.NArgs, V) :-
variabilise_last(Arg2.Args, Arg, NArgs, V).
match_probability([p(V0=C)=Prob|_], C, V, Prob) :-
V0 == V,
!.
match_probability([_|Probs], C, V, Prob) :-
match_probability(Probs, C, V, Prob).
:- use_parfactors(on) -> true ; assert(use_parfactors(off)).