-
Notifications
You must be signed in to change notification settings - Fork 2
/
rules.py
627 lines (552 loc) · 23.7 KB
/
rules.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
try:
from opencog.atomspace import TruthValue, confidence_to_count, types as t
except ImportError:
from atomspace_remote import TruthValue, types as t, confidence_to_count, types as t
import formulas
from tree import *
import math
def evaluation_link_template(predicate = None, arguments = None):
if predicate == None:
predicate = new_var()
if arguments == None:
args_tree = new_var()
else:
args_tree = T('ListLink', arguments)
return T('EvaluationLink',
predicate,
args_tree)
def actionDone_template(atomspace):
return evaluation_link_template(
atomspace.add(t.PredicateNode, name='actionDone'),
[new_var()]
)
def rules(a, deduction_types):
'''This function sets up all the Rules available in PLN. The concept of a Rule is the same as
in higher-order logic. For example, Modus Ponens requires an ImplicationLink $1->$2 as well as the
premise of the ImplicationLink ($1) and will produce $2. Every Rule comes with its own formula
for calculating TruthValues. PLN works by applying a series of Rules to some Atoms to create new
ones.'''
# This function creates a list of Rules using the Rule constructor.
# See the Rule class for more explanation of the arguments used.
# In every step of chaining, the backward chainer has a target Atom (usually with some variables)
# it is trying to find. It will apply any Rule whose head unifies with the target.
# The function is called every time a Chainer is constructed.
rules = []
#path_rules(a)
# PLN is able to do pathfinding (experimental). In the Minecraft world, there is a 3D grid
# similar to that used by A*. This Rule determines the grid squares (or rather grid cubes)
# that are next to a specified grid cube (i.e. the neighborhood, in the A* sense).
# You can represent the grid as a network of OpenCog links, which is more suitable for PLN
# than having an array (and more versatile in general).
# This version of the Rule calculates the neighborhood dynamically. The function path_rules()
# uses the simpler but much slower approach of adding thousands of links to the AtomSpace in advance.
r = Rule(T('EvaluationLink',
a.add(t.PredicateNode,'neighbor'),
T('ListLink',
Var(1), # a ConceptNode representing the coordinates in a tacky format
Var(2)
)),
[],
name='Neighbors',
match=match_neighbors)
rules.append(r)
#r = Rule(T('ImplicationLink',
# Var(1),
# Var(2)
# ),
# [],
# name='Neighbors',
# match=match_neighbors)
#rules.append(r)
# You can add a separate Rule for every axiom (i.e. every Atom which has a TruthValue when PLN starts).
# It won't add a new Rule for atoms that are created during the inference, rather they will be added to
# the Proof DAG. It's simple to have a separate Rule for each axiom, but it's slow because the chainer
# will try to unify a target against every axiom Rule.
#for obj in a.get_atoms_by_type(t.Atom):
# # POLICY: Ignore all false things. This means you can never disprove something! But much more useful for planning!
# if obj.tv.count > 0 and obj.tv.mean > 0:
# tr = tree_from_atom(obj)
# # A variable with a TV could just prove anything; that's evil!
# if not tr.is_variable():
#
# # tacky filter
# if 'CHUNK' in str(tr):
# continue
#
# r = Rule(tr, [], '[axiom]', tv = obj.tv)
# rules.append(r)
# Just lookup the rule rather than having separate rules. Would be faster
# with a large number of atoms (i.e. more scalable). Some examples will break if
# you use it due to bugs in the backward chainer.
r = Rule(Var(123),[],
name='Lookup',
match=match_axiom_slow)
rules.append(r)
# A simple example Rule to test the mechanism. You're allowed to add a Rule which calls
# any Python function to decide one of the Atoms. This one does the plus calculation.
# The first two variables are the numbers to be added, and the third variable is the result.
# i.e. the last variable is the return-value for the '+' function. This is a common pattern in
# Prolog.
r = Rule(T('EvaluationLink',
a.add(t.PredicateNode,'+'),
T('ListLink',
Var(1),
Var(2),
Var(3))),
[],
name='PredicateEvaluation',
match=match_predicate)
rules.append(r)
# The three main logical rules in PLN. The code creates different versions of the Rule
# for different kinds of Links.
# The classic Modus Ponens rule, used (and over-emphasized) in classical logic everywhere.
for ty in ['ImplicationLink', 'PredictiveImplicationLink']:
rules.append(Rule(Var(2),
[T(ty, 1, 2),
Var(1) ],
name='ModusPonens '+ty,
formula = formulas.modusPonensFormula))
# The PLN DeductionRule. Not to be confused with ModusPonens.
for type in deduction_types:
rules.append(Rule(T(type, 1,3),
[T(type, 1, 2),
T(type, 2, 3),
Var(1),
Var(2),
Var(3)],
name='Deduction',
formula = formulas.deductionSimpleFormula))
# PLN InversionRule, which reverses an ImplicationLink. It's based on Bayes' Theorem.
for type in deduction_types:
rules.append(Rule( T(type, 2, 1),
[T(type, 1, 2),
Var(1),
Var(2)],
name='Inversion',
formula = formulas.inversionFormula))
# Calculating logical And/Or/Not. These have probabilities attached,
# but they work similarly to the Boolean versions.
type = 'AndLink'
for size in xrange(5):
args = [new_var() for i in xrange(size+1)]
rules.append(Rule(T(type, args),
args,
type[:-4],
formula = formulas.andSymmetricFormula))
type = 'OrLink'
for size in xrange(1,2):
args = [new_var() for i in xrange(size+1)]
rules.append(Rule(T(type, args),
args,
type[:-4],
formula = formulas.orFormula))
# Calculate (NotLink A) using A. There's currently no Rule to find A using (NotLink A)
rules.append(Rule(T('NotLink', 1),
[ Var(1) ],
name = 'Not',
formula = formulas.notFormula))
# PLN's heuristic Rules to convert one kind of link to another. There are other
# variations on this Rule defined in the PLN book, but not implemented yet.
rules.append(Rule(T('InheritanceLink', 1, 2),
[ T('SubsetLink', 1, 2) ],
name = 'SubsetLink=>InheritanceLink',
formula = formulas.ext2InhFormula))
# Used by planning. An ExecutionLink indicates an action being performed, so we can
# assume that the action will be performed as part of the plan (i.e. if any action
# occurs in the plan, set the TV to full probability and confidence).
#r = Rule(T('ExecutionLink', 1, 2),
# [],
# name = 'PerformAction',
# tv = TruthValue(1.0,confidence_to_count(1.0)))
#rules.append(r)
# TOdo this should be actionSuccess not just actionDone
r = Rule(actionDone_template(a),
[],
name = 'PerformAction',
tv = TruthValue(1.0,confidence_to_count(1.0)))
rules.append(r)
# If something is true at the current time, you can use it as the start of the plan.
# First find the latest time (hack)
current_time = 0
for time in a.get_atoms_by_type(t.TimeNode):
timestamp = int(time.name)
if timestamp > current_time:
current_time = timestamp
current_time_node = a.add(t.TimeNode, str(current_time))
# Then create the Rule
# It's essential to store the template so it will have the same variables in both
# the head and the goal
template = evaluation_link_template()
r = Rule(template,
[T('AtTimeLink', current_time_node, template)],
name = 'AtCurrentTime'
)
rules.append(r)
# # Producing ForAll/Bind/AverageLinks.
# for type in ['ForAllLink', 'BindLink', 'AverageLink']:
# rules.append(Rule(T(type, 1, 2),
# [ Var(2) ],
# name = type+' abstraction',
# formula = formulas.identityFormula))
# This may cause weirdness with things matching too eagerly...
# # Both of these rely on the policy that tree_from_atom replaces VariableNodes in the AtomSpace with the variables the tree class uses.
# fact = new_var()
# list_link = new_var()
# r = Rule(
# fact,
# [T('ForAllLink', list_link, fact )],
# name = 'ForAll'
# )
# r.tv = True
# rules.append(r)
# If an Atom is available in an Average/ForAll quantifier, you want to be
# able to produce the Atom itself and send it through the other steps of the
# inference.
for atom in a.get_atoms_by_type(t.AverageLink):
# out[0] is the ListLink of VariableNodes, out[1] is the expression
tr = tree_from_atom(atom.out[1])
r = Rule(tr, [], name='Average')
r.tv = atom.tv
rules.append(r)
for atom in a.get_atoms_by_type(t.ForAllLink):
# out[0] is the ListLink of VariableNodes, out[1] is the expression
tr = tree_from_atom(atom.out[1])
r = Rule(tr, [], name='ForAll')
r.tv = atom.tv
rules.append(r)
rules += temporal_rules(a)
# Return every Rule specified above.
return rules
def temporal_rules(atomspace):
rules = []
rules.append(Rule(T('BeforeLink', 1, 2),
[ ],
name='Before',
match = create_temporal_matching_function(formulas.beforeFormula)
)
)
return rules
def lookup_times(tree, atomspace):
'''For the specified Tree (which can contain a Node or Link), search for
all the AtTimeLinks. Return a dictionary from timestamp (as a string) to
a float, representing the fuzzy truth value. (Ignore confidence which is
probably OK.)'''
template = T('AtTimeLink',
Var(0),
tree)
attimes = find(template, atomspace.get_atoms_by_type(t.AtTimeLink))
dist = {}
for link in attimes:
assert isinstance(link, Atom)
time = link.out[0].name
fuzzy_tv = link.out[1]
dist[time] = fuzzy_tv
return dist
def create_temporal_matching_function(formula):
def match_temporal_relationship(space,target):
assert isinstance(target, Tree)
assert not target.args[0].is_variable()
assert not target.args[1].is_variable()
distribution_event1 = lookup_times(target.args[0], space)
distribution_event2 = lookup_times(target.args[1], space)
strength = formula(distribution_event2, distribution_event2)
# I'm not sure what to choose for this
count = 1
tv = TruthValue(strength, count)
return [(target,tv)]
return match_temporal_relationship
#def path_rules(a):
# # These are some tacky and inefficient (but conceptually interesting) pathfinding experiments.
# template = T('LatestLink',
# T('AtTimeLink',
# Var(-1),
# T('EvaluationLink',
# T(a.add(t.PredicateNode, 'AGISIM_position')),
# T('ListLink',
# Var(0),
# Var(1),
# Var(2),
# Var(3)
# )
# )
# )
# )
#
# #print template
#
# def change_position(coords):
# s = {Var(i+1):T(a.add(t.NumberNode, str(c))) for (i, c) in enumerate(coords)}
# return subst(s, template)
#
# def get_position(tr):
# s = unify(template, tr, {})
# return tuple(float(s[Var(i)].op.name) for i in [1,2,3])
#
# def get_name(tr):
# s = unify(template, tr, {})
# return s[Var(0)].op.name
#
# def is_block(tr):
# return get_name(tr).startswith('id_CHUNK')
#
# lls = a.get_atoms_by_type(t.LatestLink)
# pls = find_tree(template,lls)
# #print len(pls)
#
# taken_grids = set()
# # Find any objects/blocks and assume they are obstacles.
# for pl in pls:
# #if is_block(pl):
# # continue
# (x,y,z) = get_position(pl)
# #print (x,y,z), get_name(pl)
# (x,y,z) = tuple(map(math.floor,(x,y,z)))
# #print (x,y,z), get_name(pl)
# taken_grids.add((x,y,z))
#
# # add in the ground blocks.
# ground = []
# z = 98
# for x in xrange(0,100):
# for y in xrange(0,100):
# #p1 = change_position((x,y,z))
# #ground.append(p1)
# p1c = (x,y,z)
#
# if (x,y,z+1) in taken_grids:
# #print 'occupied'
# continue
#
# ns = [(x-1,y,z),(x+1,y,z),(x,y-1,z),(x,y+1,z)] #,(x,y,z-1),(x,y,z+1)]
# ns += [(x,y,z+1) for n in ns]
# for n in ns:
# #p2 = change_position((i,j,z))
#
# #r = T('ImplicationLink',
# # a.add(t.ConceptNode,'at %s' % str(p1c)),
# # a.add(t.ConceptNode,'at %s' % str(n))
# #)
# r = T('EvaluationLink',
# a.add(t.PredicateNode,'neighbor'),
# T('ListLink',
# a.add(t.ConceptNode,'at %s' % str(p1c)),
# a.add(t.ConceptNode,'at %s' % str(n))
# )
# )
# ra = atom_from_tree(r, a)
# ra.tv = TruthValue(1,confidence_to_count(1.0))
# #print ra
#
# # add in any other surfaces (made of blocks)
# for pl in pls:
# if not is_block(pl):
# continue
#
# coords = get_position(pl)
# #print coords
# (x,y,z) = coords
# (x,y,z) = (x-0.5,y-0.5,z)
#
# if (x,y,z+1) in taken_grids:
# print 'occupied'
# continue
#
# ns = [(x-1,y,z),(x+1,y,z),(x,y-1,z),(x,y+1,z)]
# ns += [(x,y,z+1) for n in ns]
# #print pl
# for n in ns:
# neighbor_block_template = change_position(n)
# #print neighbor_block_template
# tmp = find_tree(neighbor_block_template, pls)
# # Pray there is nothing else at that exact coordinate
# assert len(tmp) < 2
#
# if len(tmp):
# # Can go straight from one block to another
# #tr = T('ImplicationLink',
# # a.add(t.ConceptNode,'at %s' % str(coords)),
# # a.add(t.ConceptNode,'at %s' % str(n))
# #)
# tr = T('EvaluationLink',
# a.add(t.PredicateNode,'neighbor'),
# T('ListLink',
# a.add(t.ConceptNode,'at %s' % str(coords)),
# a.add(t.ConceptNode,'at %s' % str(n))
# )
# )
# #print tr
# atom_from_tree(tr, a).tv = TruthValue(1,confidence_to_count(1.0))
#
# return []
#
# #for obj in a.get_atoms_by_type(t.ObjectNode):
# # if not obj.name.startswith('id_CHUNK'):
# # continue
# #
# # block = T(obj)
# See how they are used in rules() above.
# The slow one but still used because match_axiom still has bugs
def match_axiom_slow(space,target):
if isinstance(target.op, Atom):
candidates = [target.op]
else:
candidates = space.get_atoms_by_type(target.get_type())
candidates = [c for c in candidates if c.tv.count > 0]
candidate_trees = (tree_from_atom(atom) for atom in candidates)
candidate_tvs = (c.tv for c in candidates)
return zip(candidate_trees, candidate_tvs)
def match_axiom(space,target):
if isinstance(target.op, Atom):
candidates = [target.op]
else:
# The nodes in the target, as Atoms
nodes = [tr.op for tr in target.flatten() if isinstance(tr.op, Atom) and tr.op.is_node()]
if not len(nodes):
return match_axiom_slow(target)
# TODO choose the node with the smallest incoming set.
n = nodes[0]
candidates = find_links_upward(n)
candidate_trees = (tree_from_atom(atom) for atom in candidates)
candidate_tvs = (c.tv for c in candidates)
return zip(candidate_trees, candidate_tvs)
def find_links_upward(atom):
level = [link for link in atom.incoming if link.tv.count > 0]
next_level = concat_lists([find_links_upward(link) for link in atom.incoming])
return level + next_level
def match_neighbors(space,target):
print 'match_neighbors',target
template = T('EvaluationLink',
space.add(t.PredicateNode,'neighbor'),
T('ListLink',
Var(1),
Var(2)
)
)
s = unify(template, target, {})
cube1 = s[Var(1)]
#template = T('ImplicationLink',
# Var(2),
# Var(1)
#)
#s = unify(template, target, {})
#cube1 = s[Var(1)]
def get_coords(expr):
name = expr.op.name
tuple_str = name[3:]
coords = tuple(map(int, tuple_str[1:-1].split(',')))
return coords
if cube1.is_variable() or not isinstance(cube1.op, Atom):
return []
(x,y,z) = get_coords(cube1)
ns = [(x-1,y,z),(x+1,y,z),(x,y-1,z),(x,y+1,z)]
# jumping onto a block
ns += [(x,y,z+1) for n in ns]
tv = TruthValue(1,confidence_to_count(1.0))
candidates = []
for n in ns:
cube2 = Tree(space.add(t.ConceptNode, 'at %s' % str(n)))
s = {Var(1):cube1, Var(2):cube2}
cand = subst(s, template)
candidates.append((cand,tv))
return candidates
def match_predicate(space,target):
# Actually it would be much better to use the standard target for this rule
# + unify to find the places
#print 'match_predicate',target
ll = target.args[1]
#print ll
(n1,n2,n3) = ll.args
candidates = []
if (n1.get_type() == t.NumberNode and n2.get_type() == t.NumberNode):
if (n3.get_type() == t.NumberNode):
if int(n1.op.name) + int(n2.op.name) == int(n3.op.name):
candidates.append((target, TruthValue(1.0,confidence_to_count(1.0))))
elif (n3.is_variable()):
addition = int(n1.op.name) + int(n2.op.name)
s = {n3: Tree(space.add(t.NumberNode,str(addition)))}
c = subst(s, target)
print '>>>match_predicate:c',c
candidates.append((c, None))
return candidates
## @}
class Rule :
def __init__ (self, head, goals, name, tv = TruthValue(0, 0),
formula = None, match = None):
'''@head is a Tree representing the structure of Atom (usually a Link)
to be produced by this Rule. If it's a variable then any kind of Atom
can be produced.
@goals (list of Trees) specifies what kinds of Atoms
are necessary to produce it. There should be variables used in the head
that also appear in the goals. This means they will be the same atom.
It's OK to reuse the same variable numbers in different Rules - they'll
be converted to unique variables automatically.
The @name of the Rule is just for logging/visualizing the inference process.
You can use either @formula or @tv. If you specify a formula method from formulas.py
it will be called with the TVs of the relevant Atoms, to calculate the TruthValue for
the resulting Atom. If you set the tv parameter of the Rule, it will always use that TV.
@match lets you use custom code to decide which atoms are produced (or used). The API for that
is still in flux so not documented.'''
self.head = head
self.goals = goals
self.name = name
self.tv = tv
self.match = match
self.formula = if_(formula, formula, formulas.identityFormula)
# Dingjie's tracing experiments
self.path_pre = None
self.path_axiom = None
if name == 'Lookup':
assert len(goals) == 0
#self.bc_depth = 0
def __str__(self):
#rep = self.name + ' ' + str(self.head) + ' ' + str(self.tv)
rep = self.name + ' ' + str(self.head)
rep += ' :- '
for goal in self.goals :
#rep += ' '*(self.bc_depth*3+3)
rep += str(goal) + ','
return rep
def __repr__ (self) :
rep = self.name + ' ' + str(self.head) + ' ' + str(self.tv)
rep += ' :- '
for goal in self.goals :
#rep += ' '*(self.bc_depth*3+3)
rep += str(goal) + ','
return rep
def standardize_apart(self):
'''Create a new version of the Rule where both the head and goals use new
variables. Important for unification.'''
head_goals = (self.head,)+tuple(self.goals)
tmp = standardize_apart(head_goals)
new_version = Rule(tmp[0], tmp[1:], name=self.name, tv = self.tv,
formula=self.formula, match = self.match)
return new_version
def isomorphic(self, other):
'''Check if both self and other are the same. Renaming a variable doesn't affect it.
It's possible for the same target to be produced multiple times by the backward chainer,
so this enables you to reuse the existing one.'''
# One way: make conjunctions out of the rules to make
# sure variable renamings are consistent across both
# conclusion and premises
self_conj = (self.head,)+tuple(self.goals)
other_conj = (other.head,)+tuple(other.goals)
return isomorphic_conjunctions_ordered(self_conj, other_conj)
def canonical_tuple(self):
try:
return self._tuple
except:
conj = (self.head,)+tuple(self.goals)
self._tuple = tuple(canonical_trees(conj))
return self._tuple
def unifies(self, other):
'''return true if could be unified to tree other '''
self_conj = (self.head,)+tuple(self.goals)
other_conj = (other.head,)+tuple(other.goals)
return unify(self_conj, other_conj, {}) != None
def subst(self, s):
'''Create a new Rule where substitution s has been substituted into both the head and goals.'''
new_head = subst(s, self.head)
new_goals = list(subst_conjunction(s, self.goals))
new_rule = Rule(new_head, new_goals, name=self.name, tv = self.tv,
formula = self.formula, match = self.match)
return new_rule