def play_episode(root, max_depth, apprentice, move_type="all", verbose=False):
episode = []
state = copy.deepcopy(root)
edge_index = boardToData(root).edge_index
# ******************* PLAY EPISODE ***************************
for i in range(max_depth):
#print_message_over(f"Playing episode: {i}/{max_depth}")
# Check if episode is over
if state.gameOver: break
# Check is current player is alive or not
if not state.activePlayer.is_alive:
# print("\npassing, dead player")
state.endTurn()
continue
# Get possible moves, and apprentice policy
mask, actions = agent.maskAndMoves(state, state.gamePhase, edge_index)
try:
policy, value = apprentice.getPolicy(state)
except Exception as e:
state.report()
print(state.activePlayer.is_alive)
print(state.activePlayer.num_countries)
raise e
if isinstance(mask, torch.Tensor):
mask = mask.detach().numpy()
probs = policy * mask
probs = probs.flatten()
probs = probs / probs.sum()
# Random selection? e-greedy?
ind = np.random.choice(range(len(actions)), p=probs)
move = agent.buildMove(state, actions[ind])
saved = (move_type == "all" or move_type == state.gamePhase)
if verbose:
# print(f"\t\tPlay episode: turn {i}, move = {move}, saved = {saved}")
pass
if saved:
episode.append(copy.deepcopy(state))
# Play the move to continue
state.playMove(move)
return episode
def play_episode(root, max_depth, apprentice):
episode = []
state = copy.deepcopy(root)
edge_index = boardToData(root).edge_index
# ******************* PLAY EPISODE ***************************
for i in range(max_depth):
#print_message_over(f"Playing episode: {i}/{max_depth}")
# Check if episode is over
if state.gameOver: break
# Check is current player is alive or not
if not state.activePlayer.is_alive:
# print("\npassing, dead player")
state.endTurn()
continue
# Get possible moves, and apprentice policy
mask, actions = maskAndMoves(state, state.gamePhase, edge_index)
try:
policy, value = apprentice.play(state)
except Exception as e:
state.report()
print(state.activePlayer.is_alive)
print(state.activePlayer.num_countries)
raise e
policy = policy * mask
probs = policy.squeeze().detach().numpy()
probs = probs / probs.sum()
# Random selection? e-greedy?
ind = np.random.choice(range(len(actions)), p = probs)
move = buildMove(state, actions[ind])
episode.append(copy.deepcopy(state))
# Play the move to continue
state.playMove(move)
return episode
#%%%
board.play()
while not board.gameOver and board.gamePhase != "attack":
board.play()
board.report()
print(board.countriesPandas())
print("\n")
# Get policy for board
canon, _ = board.toCanonical(board.activePlayer.code)
batch = torch_geometric.data.Batch.from_data_list([boardToData(canon)])
mask, moves = agent.maskAndMoves(canon, canon.gamePhase, batch.edge_index)
policy, value = apprentice.getPolicy(canon)
pop = policy.squeeze()
T = 1
exp = np.exp(np.log(np.maximum(pop, 0.000001))/T)
soft = exp/exp.sum()
co = board.countries()
for m, a, p, s in zip(mask.squeeze(), moves, pop, soft):
if m.item():
if len(a) > 2:
print(
f"{a[0]}: {co[a[1]]['id']} -> {co[a[2]]['id']} - {p:.3f} - {s:.3f}")
else:
print(f"{a[0]}: {co[a[1]]['id']} - {p:.3f}- {s:.3f}")