def __init__(self, teacher_iters, QM, e, epsilon):
self.eps = epsilon
self.placement = None
self.penum = PlacementEnumerator(e)
self.teacher_iters = teacher_iters
self.teacher = LowestCenterOfGravityAgent(e)
self.QM = QM
class FullPlacementActor(object):
def __init__(self, teacher_iters, QM, e, epsilon):
self.eps = epsilon
self.placement = None
self.penum = PlacementEnumerator(e)
self.teacher_iters = teacher_iters
self.teacher = LowestCenterOfGravityAgent(e)
self.QM = QM
def act(self, s, debug_mode=False):
debug_info = {}
if s.t < self.teacher_iters:
debug_info["pfbms"] = []
return (self.teacher.act(s), debug_info)
if self.placement is None or not self.placement.is_incomplete():
choices = self.penum.get_placement_actions(s)
debug_info["pfbms"] = map(lambda x: x.pfbm, choices)
if random.random() < self.eps:
p = random.choice(choices)
else:
pass
#all_actions = map(lambda x: x.minor_actions,choices)
#all_bitmaps = map(lambda x: x.final_state.arena.bitmap[-1].flatten(),choices)
#initial_bitmaps = [s.arena.bitmap[-1].flatten() for i in all_bitmaps]
#print "{} choices. Bitmaps/Bitmaps/Betas: {}".format(len(choices), list(zip(initial_bitmaps,all_bitmaps,map(lambda x: self.QM.beta(s,x), choices))))
#raise Exception("TODO")
p = max(choices, key=lambda x: self.QM.q(s, x))
self.placement = PartiallyCompletedPlacement(p)
else:
debug_info["pfbms"] = []
return (self.placement.incr_action(), debug_info)
class LowestCenterOfGravityAgent(object):
penum = None
queued_actions = None
e = None
def __init__(self, e):
self.penum = PlacementEnumerator(e)
self.queued_actions = []
def act(self, s, debug_mode=False):
def score(b):
row_indices = np.nonzero(b)[0]
return -row_indices.mean()
if len(self.queued_actions) > 0:
a = self.queued_actions[0]
self.queued_actions = self.queued_actions[1:]
return (a, None)
else:
logging.info("DELTA: 0")
a_s_p = self.penum._get_actionseq_finalstate_pairs(s)
assert (len(a_s_p) > 0)
amin, smin, pmin = min(a_s_p, key=lambda x: score(x[2]))
#assert pmin.sum()==0 or np.nonzero(pmin)[0].min() == pmin.shape[0]-1
self.queued_actions = amin[1:]
return (amin[0], None)
def observe_sars_tuple(self, s, a, r, sprime, pfbm=None):
logging.info("REWARD: {}".format(r))
pass
def save_model(self, fn):
pass
class LowestCenterOfGravityAgent(object):
penum=None
queued_actions=None
e=None
def __init__(self,e):
self.penum=PlacementEnumerator(e)
self.queued_actions=[]
def act(self,s,debug_mode=False):
def score(b):
row_indices = np.nonzero(b)[0]
return -row_indices.mean()
if len(self.queued_actions)>0:
a = self.queued_actions[0]
self.queued_actions = self.queued_actions[1:]
return (a,None)
else:
logging.info("DELTA: 0")
a_s_p = self.penum._get_actionseq_finalstate_pairs(s)
assert(len(a_s_p)>0)
amin,smin,pmin = min(a_s_p, key=lambda x:score(x[2]))
#assert pmin.sum()==0 or np.nonzero(pmin)[0].min() == pmin.shape[0]-1
self.queued_actions = amin[1:]
return (amin[0], None)
def observe_sars_tuple(self,s,a,r,sprime,pfbm=None):
logging.info("REWARD: {}".format(r))
pass
def save_model(self,fn):
pass
def __init__(self, teacher_iters, QM, e, epsilon):
self.eps = epsilon
self.placement = None
self.penum = PlacementEnumerator(e)
self.teacher_iters = teacher_iters
self.teacher = LowestCenterOfGravityAgent(e)
self.QM = QM
class FullPlacementActor(object):
def __init__(self, teacher_iters, QM, e, epsilon):
self.eps = epsilon
self.placement = None
self.penum = PlacementEnumerator(e)
self.teacher_iters = teacher_iters
self.teacher = LowestCenterOfGravityAgent(e)
self.QM = QM
def act(self, s, debug_mode=False):
debug_info = {}
if s.t < self.teacher_iters:
debug_info["pfbms"] = []
return (self.teacher.act(s), debug_info)
if self.placement is None or not self.placement.is_incomplete():
choices = self.penum.get_placement_actions(s)
debug_info["pfbms"] = map(lambda x: x.pfbm, choices)
if random.random() < self.eps:
p = random.choice(choices)
else:
pass
# all_actions = map(lambda x: x.minor_actions,choices)
# all_bitmaps = map(lambda x: x.final_state.arena.bitmap[-1].flatten(),choices)
# initial_bitmaps = [s.arena.bitmap[-1].flatten() for i in all_bitmaps]
# print "{} choices. Bitmaps/Bitmaps/Betas: {}".format(len(choices), list(zip(initial_bitmaps,all_bitmaps,map(lambda x: self.QM.beta(s,x), choices))))
# raise Exception("TODO")
p = max(choices, key=lambda x: self.QM.q(s, x))
self.placement = PartiallyCompletedPlacement(p)
else:
debug_info["pfbms"] = []
return (self.placement.incr_action(), debug_info)
class LowestCenterOfGravityAgent(object):
penum=None
queued_actions=None
e=None
def __init__(self,e):
self.penum=PlacementEnumerator()
self.queued_actions=[]
self.e = e
def act(self,s):
def score(b):
row_indices = np.nonzero(b)[0]
return -row_indices.mean()
if len(self.queued_actions)>0:
a = self.queued_actions[0]
self.queued_actions = self.queued_actions[1:]
return a
else:
bts = self.penum.get_successor_bitmaps(self.e, s)
assert(len(bts)>0)
smin,amin,rmin = min(bts, key=lambda x:score(x[0].arena.bitmap))
self.queued_actions = amin[1:]
return amin[0]
def __init__(self,e):
self.penum=PlacementEnumerator(e)
self.queued_actions=[]
def __init__(self, e):
self.penum = PlacementEnumerator(e)
self.queued_actions = []
