def evaluate(index):
card = cards[index]
choices = stacks[np.array([s[-1] < card for s in stacks])]
if len(choices) == 0:
e = (1.0 - card / NUM_CARDS) * min(
count_nimmts(s) for s in stacks)
else:
stack = choices[np.argmax([s[-1] for s in choices])]
if len(stack) == STACK_VOL:
e = (1.0 -
(card - stack[-1]) / NUM_CARDS) * count_nimmts(stack)
else:
e = ((-0.1 + card - stack[-1]) / NUM_CARDS) / (STACK_VOL -
len(stack))
next = cards[index +
1] if index < len(cards) - 1 else NUM_CARDS
e += ((len(stack) == STACK_VOL - 1) * 2 -
1) * (1.0 - (next - card)) / NUM_CARDS
return e
def evaluate(index):
card = cards[index]
choices = stacks[np.array([s[-1] < card for s in stacks])]
if len(choices) == 0:
e = (1.0 - card / NUM_CARDS) * min(
count_nimmts(s) for s in stacks)
else:
stack = choices[np.argmax([s[-1] for s in choices])]
if len(stack) == STACK_VOL:
e = (1.0 -
(card - stack[-1]) / NUM_CARDS) * count_nimmts(stack)
else:
e = ((-0.1 + card - stack[-1]) / NUM_CARDS) / (
STACK_VOL - len(stack)) * count_nimmts(stack)
next = NUM_CARDS
for s in stacks:
if s[-1] > card:
next = min(next, s[-1])
#if len(stack) =4:
# e -= (next - card)
if card != cards[-1]:
next = min(next, cards[index + 1])
if len(stack) == 4:
e -= (next - card -
1.0) / NUM_CARDS * count_nimmts(stack + [
card,
])
else:
e += (next - card -
1.0) / NUM_CARDS * count_nimmts(stack + [
card,
])
#else:
# e -=
#e += ((len(stack) == STACK_VOL - 1) * 2 - 1) * (1.0 - (next - card)) / NUM_CARDS
return e
def normalize(dat) :
bits = [];
action = dat['action']
card_stacks = dat['state']['card_stacks']
card_status = dat['state']['card_status']
agent_id = dat['state']['agent_id']
hand_card = dat['state']['hand_cards'][agent_id]
for i in range(16):
l_and = True
l_or = False
for j in range(4):
if ((1<<j)&i > 0):
l_and = l_and and (action < card_stacks[j][-1])
l_or = l_or or (action < card_stacks[j][-1])
bits.append(int(l_and))
bits.append(int(l_or))
bits.append(1.0 * action / pyrl.common.num_agent_init_card)
for i in range(4):
bits.append(1.0 * card_stacks[i][-1] / pyrl.common.num_agent_init_card)
tmp = []
for i in range(4):
for j in range(5):
tmp.append(j == len(card_stacks[i]))
target_stack = -1
for j in range(4):
if action > card_stacks[j][-1]:
if target_stack == -1 or card_stacks[j][-1] > card_stacks[target_stack][-1]:
target_stack = j
for i in range(4):
tmp.append(action > card_stacks[i][-1])
tmp.append(target_stack == i)
tmp.append(target_stack == -1)
for i in range(num_base_ind):
bits.append(int(tmp[i]))
for i in range(20):
for j in range(20, num_base_ind):
bits.append(int(tmp[i] and tmp[j]))
i2 = ind_2
i3 = ind_3
i4 = ind_3
for i in range(50):
bits.append(int(tmp[i2[i][0]] and tmp[i2[i][1]]))
bits.append(int(tmp[i3[i][0]] and tmp[i3[i][1]] and tmp[i3[i][2]]))
bits.append(int(tmp[i4[i][0]] and tmp[i4[i][1]] and tmp[i4[i][2]] and tmp[i4[i][3]]))
status_bucket = [[0 for col in range(13)] for row in range(3)]
#every 13 in a bucket
for i in range(num_cards):
status_bucket[card_status[i]][i/8] += 1
bits = bits + status_bucket[0] + status_bucket[1] + status_bucket[2]
if target_stack == -1:
bits.append(min([count_nimmts(s) for s in card_stacks]))
bits.append(min([count_nimmts(s) for s in card_stacks]))
bits.append(min([count_nimmts(s) for s in card_stacks]))
else:
bits.append(count_nimmts(card_stacks[target_stack]))
bits.append(count_nimmts(card_stacks[target_stack]) * int(len(card_stacks[target_stack]) >= 5))
bits.append(count_nimmts(card_stacks[target_stack]) * int(len(card_stacks[target_stack]) >= 4))
bits.append(count_nimmts(hand_card))
punish_bucket = [[0 for col in range(7)] for row in range(3)]
for i in range(num_cards):
punish_bucket[card_status[i]][num_nimmt(i)-1] += 1
bits = bits + punish_bucket[0] + punish_bucket[1] + punish_bucket[2]
#in all there are 459 features
return bits
def policy_min(self, agentEnv):
stacks = agentEnv['card_stacks']
ret = np.argmin([count_nimmts(s) for s in stacks])
return ret