def make_encoding(ship_a, ship_b, attack_range, agent):
"""This function calculates the average time to destruction when a shoots at b.
Args:
ship_a ((Ship, str)): Attacker and hull zone tuple.
ship_b ((Ship, str)): Defender and hull zone tuple.
trials (int): Number of trials in average calculation.
range (str): Attack range.
"""
roll_counts = []
# Reset ship b for each trial
world_state = WorldState()
world_state.addShip(ship_a, 0)
world_state.addShip(ship_b, 1)
pool_colors, pool_faces = ship_a.roll("front", attack_range)
attack = AttackState(attack_range=attack_range,
attacker=ship_a,
attacking_hull="front",
defender=ship_b,
defending_hull="front",
pool_colors=pool_colors,
pool_faces=pool_faces)
world_state.updateAttack(attack)
# The defense token and die locations have been reordered in the encoding, put them back to
# their original ordering here.
encoding = Encodings.encodeAttackState(world_state)
return encoding, world_state
def a_vs_b(ship_a, ship_b, trials, attack_range):
"""This function calculates the average time to destruction when a shoots at b.
Args:
ship_a ((Ship, str)): Attacker and hull zone tuple.
ship_b ((Ship, str)): Defender and hull zone tuple.
trials (int): Number of trials in average calculation.
range (str): Attack range.
"""
roll_counts = []
agent = SimpleAgent()
for trial in range(trials):
# Reset ship b for each trial
ship_b.reset()
world_state = WorldState()
world_state.addShip(ship_a, 0)
world_state.addShip(ship_b, 1)
num_rolls = 0
while ship_b.damage_cards() < ship_b.hull():
num_rolls += 1
# Handle the attack and receive the updated world state
world_state = handleAttack(world_state=world_state,
attacker=(ship_a, "front"),
defender=(ship_b, "front"),
attack_range=attack_range,
offensive_agent=agent,
defensive_agent=agent)
roll_counts.append(num_rolls)
np_counts = numpy.array(roll_counts)
return np_counts.mean()
def a_vs_b(ship_a, ship_b, agent_a, agent_b, ship_a_hull, trials,
attack_range):
"""This function runs multiple trials of ship_a firing upon ship_b.
Args:
ship_a ((Ship, str)): Attacker and hull zone tuple.
ship_b ((Ship, str)): Defender and hull zone tuple.
agent_a (BaseAgent): Agent to control the actions of ship a.
agent_b (BaseAgent): Agent to control the actions of ship b.
ship_a_hull (str) : Attacking hull zone.
trials (int): Number of trials in average calculation.
range (str): Attack range.
Returns:
List[(str, world_state or attack effect tuple)]
"""
state_log = []
failures = 0
for _ in range(trials):
# Reset ship b for each trial
ship_b.reset()
world_state = WorldState()
world_state.addShip(ship_a, 0)
world_state.addShip(ship_b, 1)
# Begin at round 1
world_state.round = 1
# Don't attempt forever in the case of some catastrophic reoccurring error.
attempts = 0
while ship_b.damage_cards() < ship_b.hull(
) and world_state.round <= ArmadaPhases.max_rounds:
attempts += 1
# Handle the attack and receive the updated world state
#try:
world_state = handleAttack(world_state=world_state,
attacker=(ship_a, ship_a_hull),
defender=(ship_b, "front"),
attack_range=attack_range,
offensive_agent=agent_a,
defensive_agent=agent_b,
state_log=state_log)
# Record the final state with the incremented round number.
world_state.setPhase("status phase", "increment round number")
world_state.round += 1
state_log.append(('state', world_state.clone()))
#except RuntimeError as err:
# # This is fine, the random agent will do illegal things plenty of times
# pass
if 250 == attempts:
raise RuntimeError("Too many failures for ship firing simulation.")
return state_log
def a_vs_b(ship_a, ship_b, trials, attack_range):
"""This uses a random agent to choose actions during attacks from ship_a to ship_b.
Args:
ship_a ((Ship, str)): Attacker and hull zone tuple.
ship_b ((Ship, str)): Defender and hull zone tuple.
trials (int): Number of trials in average calculation.
range (str): Attack range.
Returns:
state_log (List[List[("state" or "action", (WorldState or action tuple))]])
"""
agent = RandomAgent()
state_log = []
for trial in range(trials):
# Reset ship b for each trial
ship_b.reset()
world_state = WorldState()
world_state.addShip(ship_a, 0)
world_state.addShip(ship_b, 1)
num_rolls = 0
while ship_b.damage_cards() < ship_b.hull():
num_rolls += 1
# Handle the attack and receive the updated world state
try:
world_state = handleAttack(world_state=world_state,
attacker=(ship_a, "front"),
defender=(ship_b, "front"),
attack_range=attack_range,
offensive_agent=agent,
defensive_agent=agent,
state_log=state_log)
except RuntimeError:
# This is fine, the random agent will do illegal things plenty of times
pass
return state_log
def test_defense_tokens_model(spend_defense_tokens_model):
"""Test basic network learning loop.
Test lifetime predictions during the spend defense tokens phase.
"""
network, errors, eval_errors = spend_defense_tokens_model
network.eval()
phase_name = "attack - spend defense tokens"
world_size = Encodings.calculateWorldStateSize()
action_size = Encodings.calculateActionSize(phase_name)
attack_size = Encodings.calculateAttackSize()
input_size = world_size + action_size + attack_size
target_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
batch_size = 32
batch = torch.Tensor(batch_size, input_size).to(target_device)
print("Eval errors for B were {}".format(eval_errors))
#print("First and last errors in b are {} and {}".format(eval_errors[0], eval_errors[-1]))
assert eval_errors[0] > eval_errors[-1]
# Let's examine predictions for different dice pools and spent defense tokens.
# Go through the following scenarios:
# 1.1 An attack with more than enough damage to destroy the ship
# 1.2 The same attack but a brace that would prevent destruction
# 1.3 The same attack but a redirect that would prevent destruction
# Result: 1.1 should have lower lifetime than 1.2 and 1.3
# 2.1 An attack that can barely destroy the ship
# 2.2 An attack that barely will not destroy the ship
# Result: 2.1 should have lower lifetime than 2.2.
# Ideally 1.1 and 2.1 would predict the current round.
world_state = WorldState()
world_state.round = 1
ship_a = Ship(name="Ship A", template=ship_templates["All Defense Tokens"], upgrades=[], player_number=1)
ship_b = Ship(name="Ship B", template=ship_templates["All Defense Tokens"], upgrades=[], player_number=2)
world_state.addShip(ship_a, 0)
world_state.addShip(ship_b, 1)
pool_colors, pool_faces = ['black'] * 4, ['hit_crit'] * 4
world_state.setPhase("ship phase", phase_name)
# Set the front hull zone to 2 shields
ship_b.get_range('shields')[ArmadaTypes.hull_zones.index('front')] = 2
# Set the hull to 3 (by assigning damage to reduce the remaining hull to 3)
ship_b.set('damage', ship_b.get('hull') - 3)
attack = AttackState('short', ship_a, 'left', ship_b, 'front', pool_colors, pool_faces)
world_state.updateAttack(attack)
action_encoding = torch.cat((Encodings.encodeWorldState(world_state),
Encodings.encodeAction(world_state.sub_phase, None)))
state_encoding = Encodings.encodeAttackState(world_state)
batch[0] = torch.cat(
(action_encoding.to(target_device), state_encoding.to(target_device)))
action = [("brace", (ArmadaTypes.green, None))]
action_encoding = torch.cat((Encodings.encodeWorldState(world_state),
Encodings.encodeAction(world_state.sub_phase, action)))
state_encoding = Encodings.encodeAttackState(world_state)
batch[1] = torch.cat(
(action_encoding.to(target_device), state_encoding.to(target_device)))
world_state = WorldState()
world_state.round = 1
ship_a = Ship(name="Ship A", template=ship_templates["All Defense Tokens"], upgrades=[], player_number=1)
ship_b = Ship(name="Ship B", template=ship_templates["All Defense Tokens"], upgrades=[], player_number=2)
world_state.addShip(ship_a, 0)
world_state.addShip(ship_b, 1)
pool_colors, pool_faces = ['black'] * 4, ['hit_crit'] * 2 + ['hit'] * 2
world_state.setPhase("ship phase", phase_name)
# Set the front hull zone to 2 shields
ship_b.get_range('shields')[ArmadaTypes.hull_zones.index('front')] = 2
# Set the hull to 3 (by assigning damage to reduce the remaining hull to 3)
ship_b.set('damage', ship_b.get('hull') - 3)
attack = AttackState('short', ship_a, 'left', ship_b, 'front', pool_colors, pool_faces)
world_state.updateAttack(attack)
action = [("redirect", (ArmadaTypes.green, [('left', 4)]))]
action_encoding = torch.cat((Encodings.encodeWorldState(world_state),
Encodings.encodeAction(world_state.sub_phase, action)))
state_encoding = Encodings.encodeAttackState(world_state)
batch[2] = torch.cat(
(action_encoding.to(target_device), state_encoding.to(target_device)))
round_status = network(batch[:3])
print("super cool estimated rounds of destructions are {}".format(round_status[:3]))
# Using no defense token results in destruction, the final round should be less
assert(round_status[0].item() < round_status[1].item())
assert(round_status[0].item() < round_status[2].item())
def test_resolve_attack_effects_model(resolve_attack_effects_model):
"""Test basic network learning loop.
Test lifetime predictions during the resolve attack effects phase.
"""
network, errors, eval_errors = resolve_attack_effects_model
network.eval()
phase_name = "attack - resolve attack effects"
world_size = Encodings.calculateWorldStateSize()
action_size = Encodings.calculateActionSize(phase_name)
attack_size = Encodings.calculateAttackSize()
input_size = world_size + action_size + attack_size
# First verify that errors decreased during training.
# print("Errors for A were {}".format(errors))
print("Eval errors for A were {}".format(eval_errors))
assert eval_errors[0] > eval_errors[-1]
# Let's examine predictions for different ranges and hull zones.
target_device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
batch_size = 4
batch = torch.Tensor(batch_size, input_size).to(target_device)
# Let's examine predictions for different dice pools and spent defense tokens.
# Go through the following scenarios:
# 1.1 An attack upon a ship with only 1 hull remaining
# 1.2 The same dice pool but on a ship with full hull
# 1.3 A dice pool with only blank dice
# 1.4 A dice pool with only blanks when attacking at long range.
# Create a state from resolve attack effects and an empty action.
world_state = WorldState()
world_state.round = 1
ship_a = Ship(name="Ship A", template=ship_templates["All Defense Tokens"], upgrades=[], player_number=1)
ship_b = Ship(name="Ship B", template=ship_templates["All Defense Tokens"], upgrades=[], player_number=2)
world_state.addShip(ship_a, 0)
world_state.addShip(ship_b, 1)
pool_colors, pool_faces = ['black'] * 4, ['hit_crit'] * 4
world_state.setPhase("ship phase", "attack - resolve attack effects")
ship_b.set('damage', ship_b.get('hull') - 1)
attack = AttackState('short', ship_a, 'left', ship_b, 'front', pool_colors, pool_faces)
world_state.updateAttack(attack)
action_encoding = torch.cat((Encodings.encodeWorldState(world_state),
Encodings.encodeAction(world_state.sub_phase, None)))
state_encoding = Encodings.encodeAttackState(world_state)
batch[0] = torch.cat(
(action_encoding.to(target_device), state_encoding.to(target_device)))
# Same dice pool but the defender has full hull
ship_b.set('damage', 0)
attack = AttackState('short', ship_a, 'left', ship_b, 'front', pool_colors, pool_faces)
world_state.updateAttack(attack)
action_encoding = torch.cat((Encodings.encodeWorldState(world_state),
Encodings.encodeAction(world_state.sub_phase, None)))
state_encoding = Encodings.encodeAttackState(world_state)
batch[1] = torch.cat(
(action_encoding.to(target_device), state_encoding.to(target_device)))
# Full hull and all blanks
pool_colors, pool_faces = ['black'] * 4, ['blank'] * 4
world_state.setPhase("ship phase", "attack - resolve attack effects")
attack = AttackState('short', ship_a, 'left', ship_b, 'front', pool_colors, pool_faces)
world_state.updateAttack(attack)
state_encoding = Encodings.encodeAttackState(world_state)
batch[2] = torch.cat(
(action_encoding.to(target_device), state_encoding.to(target_device)))
# Full hull, all blanks, firing at red range
pool_colors, pool_faces = ['red'] * 2, ['blank'] * 2
world_state.setPhase("ship phase", "attack - resolve attack effects")
attack = AttackState('long', ship_a, 'left', ship_b, 'front', pool_colors, pool_faces)
world_state.updateAttack(attack)
state_encoding = Encodings.encodeAttackState(world_state)
batch[3] = torch.cat(
(action_encoding.to(target_device), state_encoding.to(target_device)))
lifetime_out = network(batch)
print("super cool attack effects round estimates are {}".format(lifetime_out))
# The lifetimes should go up sequentially with the above scenarios.
# However if the ship won't be destroyed the NN can't make an accurate relative number so be
# lenient once lifetimes go above round 6. The first scenario should result in destruction
# however.
assert(lifetime_out[0].item() < 6)
for i in range(batch.size(0) - 1):
assert(lifetime_out[i].item() < lifetime_out[i+1].item() or
(lifetime_out[i].item() > 6. and lifetime_out[i+1].item() > 6.))
template=ship_templates[attacker_name],
upgrades=[],
player_number=1)
defender = ship.Ship(name=defender_name,
template=ship_templates[defender_name],
upgrades=[],
player_number=2)
# Make sure we are actually rolling dice
colors, _ = attacker.roll(attack_hull, attack_range)
if 0 < len(colors):
logging.info("{} vs {} at range {}".format(attacker_name,
defender_name,
attack_range))
world_state = WorldState()
world_state.addShip(attacker, 0)
world_state.addShip(defender, 1)
num_rolls = 0
# This will hold lists of the state action pairs for each roll
state_actions = []
while 0 < defender.hull():
num_rolls += 1
# Handle the attack and receive the updated world state
# In this initial version of the code the prediction agent won't actually take any
# actions but we need it to log the (attack_state, action) pairs
prediction_agent.rememberStateActions()
# TOOD FIXME HERE For random training sometimes the agents should be random agents.
world_state = handleAttack(world_state=world_state,
attacker=(attacker, attack_hull),
defender=(defender, defend_hull),
attack_range=attack_range,
a_colors, a_roll = ship_1.roll("front", attack_range)
if 0 < len(a_colors):
roll_counts = []
print("{} vs {} at range {}".format(ship_name_1, ship_name_2,
attack_range))
logging.info("{} vs {} at range {}".format(
ship_name_1, ship_name_2, attack_range))
for trial in range(250):
logging.info("Trial number {}".format(trial))
# Reset ship b for each trial
ship_2 = ship.Ship(name=ship_name_2,
template=ship_templates[ship_name_2],
upgrades=[],
player_number=2)
world_state = WorldState()
world_state.addShip(ship_1, 0)
world_state.addShip(ship_2, 1)
num_rolls = 0
while 0 < ship_2.hull():
num_rolls += 1
# Handle the attack and receive the updated world state
world_state = handleAttack(world_state=world_state,
attacker=(ship_1, "front"),
defender=(ship_2, "front"),
attack_range=attack_range,
offensive_agent=agent,
defensive_agent=agent)
roll_counts.append(num_rolls)
np_counts = numpy.array(roll_counts)
print(
"Ship {} destroys {} in {} average rolls, stddev = {}, at range {}."
