def train(num_episodes: int, game_interface: HFOAttackingPlayer,
features: discrete_features_v2.DiscreteFeaturesV2,
agent: QLearningAgentV4, actions: DiscreteActionsV2, reward_funct):
for ep in range(num_episodes):
# print('<Training> Episode {}/{}:'.format(ep, num_episodes))
aux_positions_names = set()
aux_actions_played = set()
while game_interface.in_game():
# Update environment features:
features.update_features(game_interface.get_state())
curr_state_id = features.get_state_index()
has_ball = features.has_ball()
# Act:
action_idx = agent.act(curr_state_id)
hfo_action: tuple = actions.map_action_idx_to_hfo_action(
agent_pos=features.get_pos_tuple(), has_ball=has_ball,
action_idx=action_idx)
# Step:
status, observation = game_interface.step(hfo_action, has_ball)
reward = reward_funct(status)
# Save metrics:
agent.save_visited_state(curr_state_id, action_idx)
agent.cum_reward += reward
aux_positions_names.add(features.get_position_name())
action_name = actions.map_action_to_str(action_idx, has_ball)
aux_actions_played.add(action_name)
# Update environment features:
prev_state_id = curr_state_id
features.update_features(observation)
curr_state_id = features.get_state_index()
agent.store_ep(state_idx=prev_state_id, action_idx=action_idx,
reward=reward, next_state_idx=curr_state_id,
has_ball=has_ball, done=not game_interface.in_game())
agent.learn()
# print(':: Episode: {}; reward: {}; epsilon: {}; positions: {}; '
# 'actions: {}'.format(ep, agent.cum_reward, agent.epsilon,
# aux_positions_names, aux_actions_played))
agent.save_metrics(agent.old_q_table, agent.q_table)
# Reset player:
agent.reset()
agent.update_hyper_parameters(episode=ep,
num_total_episodes=num_episodes)
# Game Reset
game_interface.reset()
agent.save_model()
actions_name = [actions_manager.map_action_to_str(i, has_ball=True) for i in
range(agent.num_actions)]
agent.export_metrics(training=True, actions_name=actions_name)
def test(num_episodes: int, game_interface: HFOAttackingPlayer,
features: discrete_features_v2.DiscreteFeaturesV2,
agent: QLearningAgentV5, actions: DiscreteActionsV5, reward_funct):
"""
@param num_episodes: number of episodes to run
@param game_interface: game interface, that manages interactions
between both;
@param features: features interface, from the observation array, gets the
main features for the agent;
@param agent: learning agent;
@param actions: actions interface;
@param reward_funct: reward function used
@return: (int) the avarage reward
"""
# Run training using Q-Learning
sum_score = 0
for ep in range(num_episodes):
print('<Test> {}/{}:'.format(ep, num_episodes))
# Go to origin position:
features.update_features(game_interface.get_state())
go_to_origin_position(game_interface=game_interface,
features=features,
actions=actions)
# Test loop:
while game_interface.in_game():
# Update environment features:
curr_state_id = features.get_state_index()
has_ball = features.has_ball()
# Act:
action_idx = agent.exploit_actions(curr_state_id)
hfo_action_params, num_rep = \
actions.map_action_idx_to_hfo_action(
agent_pos=features.get_pos_tuple(), has_ball=has_ball,
action_idx=action_idx)
action_name = actions.map_action_to_str(action_idx, has_ball)
# Step:
rep_counter_aux = 0
while game_interface.in_game() and rep_counter_aux < num_rep:
status, observation = game_interface.step(
hfo_action_params, has_ball)
rep_counter_aux += 1
reward = reward_funct(status)
# update features:
features.update_features(observation)
# Save metrics:
agent.save_visited_state(curr_state_id, action_idx)
sum_score += reward
# Reset player:
agent.reset(training=False)
# Game Reset
game_interface.reset()
return sum_score / num_episodes
def test(train_ep: int, num_episodes: int, game_interface: HFOAttackingPlayer,
features: discrete_features_v2.DiscreteFeaturesV2,
agent: QLearningAgentV4, actions: DiscreteActionsV2, reward_funct):
# Run training using Q-Learning
score = 0
agent.test_episodes.append(train_ep)
for ep in range(num_episodes):
print('<Test> {}/{}:'.format(ep, num_episodes))
prev_state_id =-1
while game_interface.in_game():
# Update environment features:
features.update_features(game_interface.get_state())
curr_state_id = features.get_state_index()
has_ball = features.has_ball()
# Act:
if prev_state_id != curr_state_id:
print([round(val, 2) for val in agent.q_table[curr_state_id]])
action_idx = agent.exploit_actions(curr_state_id)
hfo_action: tuple = actions.map_action_idx_to_hfo_action(
agent_pos=features.get_pos_tuple(), has_ball=has_ball,
action_idx=action_idx)
# Step:
status, observation = game_interface.step(hfo_action, has_ball)
prev_state_id = curr_state_id
# Save Metrics:
agent.save_visited_state(curr_state_id, action_idx)
agent.cum_reward += reward_funct(status)
print(':: Episode: {}; reward: {}'.format(ep, agent.cum_reward))
score += 1 if game_interface.status == GOAL else 0
# Reset player:
agent.reset(training=False)
# Game Reset
game_interface.reset()
agent.scores.append(score)
actions_name = [actions_manager.map_action_to_str(i, has_ball=True) for i in
range(agent.num_actions)]
agent.export_metrics(training=False, actions_name=actions_name)
def train(num_train_episodes: int, num_total_train_ep: int,
game_interface: HFOAttackingPlayer,
features: discrete_features_v2.DiscreteFeaturesV2,
agent: QLearningAgentV5, actions: DiscreteActionsV5,
save_metrics: bool, reward_funct):
"""
@param num_train_episodes: number of episodes to train in this iteration
@param num_total_train_ep: number total of episodes to train
@param game_interface: game interface, that manages interactions
between both;
@param features: features interface, from the observation array, gets
the main features for the agent;
@param agent: learning agent;
@param actions: actions interface;
@param save_metrics: flag, if true save the metrics;
@param reward_funct: reward function used
@return: (QLearningAgentV5) the agent
"""
for ep in range(num_train_episodes):
# Go to origin position:
features.update_features(game_interface.get_state())
go_to_origin_position(game_interface=game_interface,
features=features,
actions=actions)
# Start learning loop
aux_positions_names = set()
aux_actions_played = set()
while game_interface.in_game():
# Update environment features:
curr_state_id = features.get_state_index()
has_ball = features.has_ball()
# Act:
action_idx = agent.act(curr_state_id)
hfo_action_params, num_rep =\
actions.map_action_idx_to_hfo_action(
agent_pos=features.get_pos_tuple(), has_ball=has_ball,
action_idx=action_idx)
# Step:
rep_counter_aux = 0
while game_interface.in_game() and rep_counter_aux < num_rep:
status, observation = game_interface.step(
hfo_action_params, has_ball)
rep_counter_aux += 1
reward = reward_funct(status)
# Save metrics:
if save_metrics:
agent.save_visited_state(curr_state_id, action_idx)
agent.cum_reward += reward
aux_positions_names.add(features.get_position_name())
action_name = actions.map_action_to_str(action_idx, has_ball)
aux_actions_played.add(action_name)
# Update environment features:
prev_state_id = curr_state_id
features.update_features(observation)
curr_state_id = features.get_state_index()
agent.store_ep(state_idx=prev_state_id,
action_idx=action_idx,
reward=reward,
next_state_idx=curr_state_id,
has_ball=has_ball,
done=not game_interface.in_game())
agent.learn()
# print(':: Episode: {}; reward: {}; epsilon: {}; positions: {}; '
# 'actions: {}'.format(ep, agent.cum_reward, agent.epsilon,
# aux_positions_names, aux_actions_played))
if save_metrics:
agent.save_metrics(agent.old_q_table, agent.q_table)
# Reset player:
agent.reset()
agent.update_hyper_parameters(episode=agent.train_eps,
num_total_episodes=num_total_train_ep)
# Game Reset
game_interface.reset()
agent.save_model()
if save_metrics:
actions_name = [
actions_manager.map_action_to_str(i, has_ball=True)
for i in range(agent.num_actions)
]
agent.export_metrics(training=True, actions_name=actions_name)
return agent
def test(num_episodes: int, game_interface: HFOAttackingPlayer,
features: DiscreteFeaturesV2, agent: QLearningAgentV6,
actions: DiscreteActionsV5, reward_funct) -> float:
"""
@param num_episodes: number of episodes to run
@param game_interface: game interface, that manages interactions
between both;
@param features: features interface, from the observation array, gets the
main features for the agent;
@param agent: learning agent;
@param actions: actions interface;
@param reward_funct: reward function used
@return: (float) the average reward
"""
# Run training using Q-Learning
sum_score = 0
for ep in range(num_episodes):
# Check if server still up:
if game_interface.hfo.step() == SERVER_DOWN:
print("Server is down while testing; episode={}".format(ep))
break
# Go to origin position:
features.update_features(game_interface.get_state())
go_to_origin_position(game_interface=game_interface,
features=features,
actions=actions)
# Test loop:
debug_counter = 0 # TODO remove
while game_interface.in_game():
# Update environment features:
curr_state_id = features.get_state_index()
has_ball = features.has_ball()
# Act:
debug_counter += 1
action_idx = agent.exploit_actions(curr_state_id)
hfo_action_params, num_rep = \
actions.map_action_idx_to_hfo_action(
agent_pos=features.get_pos_tuple(), has_ball=has_ball,
action_idx=action_idx)
# Step:
status, observation = execute_action(
action_params=hfo_action_params,
repetitions=num_rep,
has_ball=has_ball,
game_interface=game_interface)
# update features:
reward = reward_funct(status)
features.update_features(observation)
sum_score += reward
if status == OUT_OF_TIME:
if debug_counter < 5:
raise NoActionPlayedError(
"agent was only able to choose {}".format(debug_counter))
# Game Reset
game_interface.reset()
print("<<TEST>> AVR reward = ", sum_score / (ep + 1))
return sum_score / num_episodes
def train(num_train_episodes: int, num_total_train_ep: int,
game_interface: HFOAttackingPlayer, features: DiscreteFeaturesV2,
agent: QLearningAgentV6, actions: DiscreteActionsV5, reward_funct):
"""
@param num_train_episodes: number of episodes to train in this iteration
@param num_total_train_ep: number total of episodes to train
@param game_interface: game interface, that manages interactions
between both;
@param features: features interface, from the observation array, gets
the main features for the agent;
@param agent: learning agent;
@param actions: actions interface;
@param reward_funct: reward function used
@return: (QLearningAgentV5) the agent
"""
sum_score = 0
sum_epsilons = 0
agent.counter_explorations = 0
agent.counter_exploitations = 0
for ep in range(num_train_episodes):
# Check if server still up:
# if game_interface.hfo.step() == SERVER_DOWN:
# raise ServerDownError("training; episode={}".format(ep))
# Go to origin position:
features.update_features(game_interface.get_state())
go_to_origin_position(game_interface=game_interface,
features=features,
actions=actions)
# Start learning loop
debug_counter = 0 # TODO remove
while game_interface.in_game():
# Update environment features:
curr_state_id = features.get_state_index()
has_ball = features.has_ball()
# Act:
debug_counter += 1
action_idx = agent.act(curr_state_id)
hfo_action_params, num_rep =\
actions.map_action_idx_to_hfo_action(
agent_pos=features.get_pos_tuple(), has_ball=has_ball,
action_idx=action_idx)
# Step:
status, observation = execute_action(
action_params=hfo_action_params,
repetitions=num_rep,
has_ball=has_ball,
game_interface=game_interface)
# Update environment features:
reward = reward_funct(status)
sum_score += reward
features.update_features(observation)
new_state_id = features.get_state_index()
agent.store_ep(state_idx=curr_state_id,
action_idx=action_idx,
reward=reward,
next_state_idx=new_state_id,
has_ball=has_ball,
done=not game_interface.in_game())
if status == OUT_OF_TIME:
if debug_counter < 5:
raise NoActionPlayedError(
"agent was only able to choose {}".format(debug_counter))
agent.learn_buffer()
agent.update_hyper_parameters(num_total_episodes=num_total_train_ep)
sum_epsilons += agent.epsilon
# Game Reset
game_interface.reset()
print("<<TRAIN>> AVR reward = ", sum_score / num_train_episodes)
print("<<TRAIN>> %Explorations={}% ".format(
round(
(agent.counter_explorations /
(agent.counter_exploitations + agent.counter_explorations)), 4) *
100))