class EarlyEagerBeamSearch(AbstractSearch):
"""A beam search that prunes early (after each state expansion) and eagerly (weeding out worse successors)"""
def __init__(self, state, beamsize, **kwargs):
assert isinstance(state, AbstractSearchState)
self.beamsize = beamsize
super(EarlyEagerBeamSearch, self).__init__(**kwargs)
self.fringe = PriorityQueue(state,
lambda x: x.score,
self.minimize,
length=0,
blockworse=False,
blockequal=False,
duplicates=kwargs['duplicates']
if 'duplicates' in kwargs else False)
self.incomplete = True
def prune(self, state):
if self.debug:
l = len(self.fringe)
print("\t[pynlpl debug] pruning with beamsize " +
str(self.beamsize) + "...",
end="",
file=stderr)
self.fringe.prunebyscore(state.score(), retainequalscore=True)
self.fringe.prune(self.beamsize)
if self.debug:
print(" (" + str(l) + " to " + str(len(self.fringe)) + " items)",
file=stderr)
def searchtop(self, n=10):
"""Return the top n best resulta (or possibly less if not enough is found)"""
solutions = PriorityQueue([],
lambda x: x.score,
self.minimize,
length=n,
blockworse=False,
blockequal=False,
duplicates=False)
for solution in self:
solutions.append(solution)
return solutions
def __init__(self, state, beamsize, **kwargs):
assert isinstance(state, AbstractSearchState)
self.beamsize = beamsize
super(EarlyEagerBeamSearch, self).__init__(**kwargs)
self.fringe = PriorityQueue(state,
lambda x: x.score,
self.minimize,
length=0,
blockworse=False,
blockequal=False,
duplicates=kwargs['duplicates']
if 'duplicates' in kwargs else False)
self.incomplete = True
class EarlyEagerBeamSearch(AbstractSearch):
"""A beam search that prunes early (after each state expansion) and eagerly (weeding out worse successors)"""
def __init__(self, state, beamsize, **kwargs):
assert isinstance(state, AbstractSearchState)
self.beamsize = beamsize
super(EarlyEagerBeamSearch,self).__init__(**kwargs)
self.fringe = PriorityQueue(state, lambda x: x.score, self.minimize, length=0, blockworse=False, blockequal=False,duplicates= kwargs['duplicates'] if 'duplicates' in kwargs else False)
self.incomplete = True
def prune(self, state):
if self.debug:
l = len(self.fringe)
print("\t[pynlpl debug] pruning with beamsize " + str(self.beamsize) + "...",end="",file=stderr)
self.fringe.prunebyscore(state.score(), retainequalscore=True)
self.fringe.prune(self.beamsize)
if self.debug: print(" (" + str(l) + " to " + str(len(self.fringe)) + " items)",file=stderr)
def __init__(self, state, **kwargs):
assert isinstance(state, AbstractSearchState)
super(HillClimbingSearch, self).__init__(**kwargs)
self.fringe = PriorityQueue([state],
lambda x: x.score,
self.minimize,
length=0,
blockworse=True,
blockequal=False,
duplicates=False)
def __init__(self, states, beamsize, **kwargs):
if isinstance(states, AbstractSearchState):
states = [states]
else:
assert all((isinstance(x, AbstractSearchState) for x in states))
self.beamsize = beamsize
if 'eager' in kwargs:
self.eager = kwargs['eager']
else:
self.eager = False
super(BeamSearch, self).__init__(**kwargs)
self.incomplete = True
self.duplicates = kwargs[
'duplicates'] if 'duplicates' in kwargs else False
self.fringe = PriorityQueue(states,
lambda x: x.score,
self.minimize,
length=0,
blockworse=False,
blockequal=False,
duplicates=self.duplicates)
def __iter__(self):
"""Generator yielding *all* valid goalstates it can find"""
i = 0
while len(self.fringe) > 0:
i += 1
if self.debug:
print("\t[pynlpl debug] *************** STARTING ROUND #" +
str(i) + " ****************",
file=stderr)
b = 0
#Create a new empty fixed-length priority queue (this implies there will be pruning if more items are offered than it can hold!)
successors = PriorityQueue([],
lambda x: x.score,
self.minimize,
length=self.beamsize,
blockworse=False,
blockequal=False,
duplicates=self.duplicates)
while len(self.fringe) > 0:
b += 1
if self.debug:
print("\t[pynlpl debug] *************** ROUND #" + str(i) +
" BEAM# " + str(b) + " ****************",
file=stderr)
#if self.debug: print >>stderr,"\t[pynlpl debug] FRINGE: ", self.fringe
state = self.poll(self.fringe)()
if self.debug:
try:
print("\t[pynlpl debug] CURRENT STATE (depth " +
str(state.depth()) + "): " + str(state),
end="",
file=stderr)
except AttributeError:
print("\t[pynlpl debug] CURRENT STATE: " + str(state),
end="",
file=stderr)
print(" hash=" + str(hash(state)), file=stderr)
try:
print(" score=" + str(state.score()), file=stderr)
except:
pass
if not self.usememory or (self.usememory and
not hash(state) in self._visited):
self.traversed += 1
#Evaluate state
if state.test(self.goalstates):
if self.debug:
print("\t[pynlpl debug] Valid goalstate, yielding",
file=stderr)
self.solutions += 1 #counts the number of solutions
yield state
elif self.debug:
print("\t[pynlpl debug] (no goalstate, not yielding)",
file=stderr)
if self.eager:
score = state.score()
#Expand the specified state and offer to the fringe
statecount = offers = 0
for j, s in enumerate(state.expand()):
statecount += 1
if self.debug >= 2:
print("\t[pynlpl debug] (Round #" + str(i) +
" Beam #" + str(b) + ") Expanded state #" +
str(j + 1) +
", offering to successor pool: " + str(s),
end="",
file=stderr)
try:
print(s.score(), end="", file=stderr)
except:
print("ERROR SCORING!", end="", file=stderr)
pass
if not self.maxdepth or s.depth() <= self.maxdepth:
if not self.eager:
#use all successors (even worse ones than the current state)
offers += 1
accepted = successors.append(s)
else:
#use only equal or better successors
if s.score() >= score:
offers += 1
accepted = successors.append(s)
else:
accepted = False
if self.debug >= 2:
if accepted:
print(" ACCEPTED", file=stderr)
else:
print(" REJECTED", file=stderr)
else:
if self.debug >= 2:
print(" REJECTED, MAXDEPTH EXCEEDED.",
file=stderr)
elif self.debug:
print(
"\t[pynlpl debug] Not offered to successor pool, maxdepth exceeded",
file=stderr)
if self.debug:
print("\t[pynlpl debug] Expanded " + str(statecount) +
" states, " + str(offers) +
" offered to successor pool",
file=stderr)
if self.keeptraversal: self._traversal.append(state)
if self.usememory: self._visited[hash(state)] = True
self.prune(
state
) #calls prune method (does nothing by default in this search!!!)
else:
if self.debug:
print(
"\t[pynlpl debug] State already visited before, not expanding again... (hash="
+ str(hash(state)) + ")",
file=stderr)
#AFTER EXPANDING ALL NODES IN THE FRINGE/BEAM:
#set fringe for next round
self.fringe = successors
#Pruning is implicit, successors was a fixed-size priority queue
if self.debug:
print("\t[pynlpl debug] (Round #" + str(i) +
") Implicitly pruned with beamsize " +
str(self.beamsize) + "...",
file=stderr)
#self.fringe.prune(self.beamsize)
if self.debug:
print(" (" + str(offers) + " to " + str(len(self.fringe)) +
" items)",
file=stderr)
if self.debug:
print("\t[pynlpl debug] Search complete: " + str(self.solutions) +
" solution(s), " + str(self.traversed) +
" states traversed in " + str(i) + " rounds with " + str(b) +
" beams",
file=stderr)
def __init__(self, state, beamsize, **kwargs):
assert isinstance(state, AbstractSearchState)
self.beamsize = beamsize
super(EarlyEagerBeamSearch,self).__init__(**kwargs)
self.fringe = PriorityQueue(state, lambda x: x.score, self.minimize, length=0, blockworse=False, blockequal=False,duplicates= kwargs['duplicates'] if 'duplicates' in kwargs else False)
self.incomplete = True
def __iter__(self):
"""Generator yielding *all* valid goalstates it can find"""
i = 0
while len(self.fringe) > 0:
i +=1
if self.debug: print("\t[pynlpl debug] *************** STARTING ROUND #" + str(i) + " ****************",file=stderr)
b = 0
#Create a new empty fixed-length priority queue (this implies there will be pruning if more items are offered than it can hold!)
successors = PriorityQueue([], lambda x: x.score, self.minimize, length=self.beamsize, blockworse=False, blockequal=False,duplicates= self.duplicates)
while len(self.fringe) > 0:
b += 1
if self.debug: print("\t[pynlpl debug] *************** ROUND #" + str(i) + " BEAM# " + str(b) + " ****************",file=stderr)
#if self.debug: print >>stderr,"\t[pynlpl debug] FRINGE: ", self.fringe
state = self.poll(self.fringe)()
if self.debug:
try:
print("\t[pynlpl debug] CURRENT STATE (depth " + str(state.depth()) + "): " + str(state),end="",file=stderr)
except AttributeError:
print("\t[pynlpl debug] CURRENT STATE: " + str(state),end="",file=stderr)
print(" hash="+str(hash(state)),file=stderr)
try:
print(" score="+str(state.score()),file=stderr)
except:
pass
if not self.usememory or (self.usememory and not hash(state) in self._visited):
self.traversed += 1
#Evaluate state
if state.test(self.goalstates):
if self.debug: print("\t[pynlpl debug] Valid goalstate, yielding",file=stderr)
self.solutions += 1 #counts the number of solutions
yield state
elif self.debug:
print("\t[pynlpl debug] (no goalstate, not yielding)",file=stderr)
if self.eager:
score = state.score()
#Expand the specified state and offer to the fringe
statecount = offers = 0
for j, s in enumerate(state.expand()):
statecount += 1
if self.debug >= 2:
print("\t[pynlpl debug] (Round #" + str(i) +" Beam #" + str(b) + ") Expanded state #" + str(j+1) + ", offering to successor pool: " + str(s),end="",file=stderr)
try:
print(s.score(),end="",file=stderr)
except:
print("ERROR SCORING!",end="",file=stderr)
pass
if not self.maxdepth or s.depth() <= self.maxdepth:
if not self.eager:
#use all successors (even worse ones than the current state)
offers += 1
accepted = successors.append(s)
else:
#use only equal or better successors
if s.score() >= score:
offers += 1
accepted = successors.append(s)
else:
accepted = False
if self.debug >= 2:
if accepted:
print(" ACCEPTED",file=stderr)
else:
print(" REJECTED",file=stderr)
else:
if self.debug >= 2:
print(" REJECTED, MAXDEPTH EXCEEDED.",file=stderr)
elif self.debug:
print("\t[pynlpl debug] Not offered to successor pool, maxdepth exceeded",file=stderr)
if self.debug:
print("\t[pynlpl debug] Expanded " + str(statecount) + " states, " + str(offers) + " offered to successor pool",file=stderr)
if self.keeptraversal: self._traversal.append(state)
if self.usememory: self._visited[hash(state)] = True
self.prune(state) #calls prune method (does nothing by default in this search!!!)
else:
if self.debug:
print("\t[pynlpl debug] State already visited before, not expanding again... (hash=" + str(hash(state)) +")",file=stderr)
#AFTER EXPANDING ALL NODES IN THE FRINGE/BEAM:
#set fringe for next round
self.fringe = successors
#Pruning is implicit, successors was a fixed-size priority queue
if self.debug:
print("\t[pynlpl debug] (Round #" + str(i) + ") Implicitly pruned with beamsize " + str(self.beamsize) + "...",file=stderr)
#self.fringe.prune(self.beamsize)
if self.debug: print(" (" + str(offers) + " to " + str(len(self.fringe)) + " items)",file=stderr)
if self.debug:
print("\t[pynlpl debug] Search complete: " + str(self.solutions) + " solution(s), " + str(self.traversed) + " states traversed in " + str(i) + " rounds with " + str(b) + " beams",file=stderr)
def searchtop(self,n=10):
"""Return the top n best resulta (or possibly less if not enough is found)"""
solutions = PriorityQueue([], lambda x: x.score, self.minimize, length=n, blockworse=False, blockequal=False,duplicates=False)
for solution in self:
solutions.append(solution)
return solutions
def test_append_maximized_fixedlength(self):
"""Fixed-length priority queue (max)"""
global values
pq = PriorityQueue(values, lambda x: x, False, 4, False, False)
result = list(iter(pq))
self.assertEqual(result, maxtomin[:4])
def test_append_minimized_blockworse(self):
"""Minimized PriorityQueue (with blockworse)"""
global values
pq = PriorityQueue(values, lambda x: x, True, 0, True, False)
result = list(iter(pq))
self.assertEqual(result, [1, 3])
def test_append_maximized_blockworse_blockequal(self):
"""Maximized PriorityQueue (with blockworse + blockequal)"""
global values
pq = PriorityQueue(values, lambda x: x, False, 0, True, True)
result = list(iter(pq))
self.assertEqual(result, [8, 6, 3])
def test_append_maximized(self):
"""Maximized PriorityQueue"""
global values
pq = PriorityQueue(values, lambda x: x, False, 0, False, False)
result = list(iter(pq))
self.assertEqual(result, maxtomin)