if __name__ == '__main__':
# use rockNum == 0 to represent a test case
rewardCandNum = 3
"""
def r1(s, a):
if s == (0, 0) and a == (1, 0): return 0.9
elif s == (0, 1) and a == (0, 1): return 0.6
else: return 0
def r2(s, a):
if s == (0, 1) and a == (1, 0): return 1
elif s == (0, 1) and a == (0, 1): return 0.6
else: return 0
def r3(s, a):
if s == (0, 0) and a == (1, 0): return 0.45
elif s == (0, 1) and a == (1, 0): return 0.5
elif s == (0, 1) and a == (0, 1): return 0.6
else: return 0
rewardSet = [r1, r2, r3]
"""
r1 = lambda s, a: s == (2, 10)
r2 = lambda s, a: s == (1, 10)
r3 = lambda s, a: s == (3, 10)
rewardSet = [r1, r2, r3]
initialPhi = [1.0 / rewardCandNum] * rewardCandNum
experiment(cmp, rewardSet, initialPhi)
puddle = [(x, y) for x in range(loc[0], loc[0] + puddleSize)
for y in range(loc[1], loc[1] + puddleSize)]
puddles.append(puddle)
for rockID in xrange(rockPerPuddle):
rocks.append(random.choice(puddle))
rewardBasic = util.Counter()
for id in xrange(len(rocks)):
rewardBasic[rocks[id]] = id + 1
# TODO use reward gen
# reward cand indicates belief on where the puddle is
rewardCandNum = puddleNum
rewards = []
for candId in xrange(rewardCandNum):
reward = rewardBasic.copy()
puddle = puddles[candId]
for loc in puddle:
reward[loc] = -10
rewards.append(reward)
initialPhi = [1.0 / len(rewards)] * len(rewards)
terminalReward = util.Counter()
for x in range(width / 3) + range(width * 2 / 3, width):
terminalReward[(x, height - 1)] = -1000
allPuddles = sum(puddles, []) # concatenate its elements (which are lists)
experiment(Domain, width, height, responseTime, horizon, rewardCandNum,
allPuddles, rewards, initialPhi, terminalReward)
return numerical
else:
return 0
return rewardFunc
rewardSet = []
"""
# a case where trajectory query (actualy state-action preference query) has worse performance than policy queries
rewardSet.append(rewardGen([((0, 0), 0), ((0, 1), 0)], 1))
rewardSet.append(rewardGen([((0, 0), 1), ((0, 1), 0)], 1))
rewardSet.append(rewardGen([((0, 0), 0), ((0, 1), 1)], 1))
rewardSet.append(rewardGen([((0, 0), 1), ((0, 1), 1)], 1))
"""
terminalReward = util.Counter()
cmp = TabularNavigationKWay(responseTime,
width,
height,
horizon=horizon,
terminalReward=terminalReward)
rocks = cmp.getStateActionPairs()
for candId in xrange(rewardCandNum):
sampledRocks = random.sample(rocks, rockNum)
rewardSet.append(rewardGen(sampledRocks, random.random()))
initialPhi = [1.0 / rewardCandNum] * rewardCandNum
experiment(cmp, rewardSet, initialPhi)
for opt, arg in opts:
if opt == '-t':
config.TRAJECTORY_LENGTH = int(arg)
elif opt == '-n':
rewardCandNum = int(arg)
elif opt == '-k':
config.NUMBER_OF_RESPONSES = int(arg)
elif opt == '-r':
random.seed(int(arg))
numpy.random.seed(int(arg))
terminalReward = util.Counter()
# three-state domain
"""
cmp = ThreeStateToy(responseTime, horizon, terminalReward)
ws = [(-1,), (1,)]
"""
# rock collection
cmp = RockCollection(responseTime, width, height, horizon, terminalReward, rockNum)
ws = [[random.random() for dim in range(config.DIMENSION)] for _ in range(rewardCandNum)]
rewardCandNum = len(ws)
initialPhi = [1.0 / rewardCandNum] * rewardCandNum
config.opts = '_'.join(map(str, [rewardCandNum, config.NUMBER_OF_RESPONSES]))
experiment(cmp, ws, initialPhi)
for _ in xrange(puddleNum):
loc = (random.randint(0, width - puddleSize - 1), random.randint(0, height - puddleSize - 1))
#loc = puddleLocs[_]
puddle = [(x, y) for x in range(loc[0], loc[0] + puddleSize) for y in range(loc[1], loc[1] + puddleSize)]
puddles.append(puddle)
for rockID in xrange(rockPerPuddle):
rocks.append(random.choice(puddle))
rewardBasic = util.Counter()
for id in xrange(len(rocks)):
rewardBasic[rocks[id]] = id + 1
# TODO use reward gen
# reward cand indicates belief on where the puddle is
rewardCandNum = puddleNum
rewards = []
for candId in xrange(rewardCandNum):
reward = rewardBasic.copy()
puddle = puddles[candId]
for loc in puddle: reward[loc] = -10
rewards.append(reward)
initialPhi = [1.0 / len(rewards)] * len(rewards)
terminalReward = util.Counter()
for x in range(width / 3) + range(width * 2 / 3, width): terminalReward[(x, height - 1)] = -1000
allPuddles = sum(puddles, []) # concatenate its elements (which are lists)
experiment(Domain, width, height, responseTime, horizon, rewardCandNum, allPuddles, rewards, initialPhi, terminalReward)
elif opt == '-n':
rewardCandNum = int(arg)
elif opt == '-k':
config.NUMBER_OF_RESPONSES = int(arg)
elif opt == '-r':
random.seed(int(arg))
numpy.random.seed(int(arg))
terminalReward = util.Counter()
# three-state domain
"""
cmp = ThreeStateToy(responseTime, horizon, terminalReward)
ws = [(-1,), (1,)]
"""
# rock collection
cmp = RockCollection(responseTime, width, height, horizon, terminalReward,
rockNum)
ws = [[random.random() for dim in range(config.DIMENSION)]
for _ in range(rewardCandNum)]
rewardCandNum = len(ws)
initialPhi = [1.0 / rewardCandNum] * rewardCandNum
config.opts = '_'.join(
map(str, [rewardCandNum, config.NUMBER_OF_RESPONSES]))
experiment(cmp, ws, initialPhi)