num_test_ep = args.num_test_ep
num_episodes = num_train_ep + num_test_ep
print("Starting Training - id={}; num_opponents={}; num_teammates={}; "
"num_episodes={};".format(agent_id, num_op, num_team, num_episodes))
# Initialize connection with the HFO server
hfo_interface = HFOAttackingPlayer(agent_id=agent_id,
num_opponents=num_op,
num_teammates=num_team)
hfo_interface.connect_to_server()
# Agent set-up
reward_function = simple_reward
features_manager = DiscreteHighLevelFeatures(num_team, num_op)
actions_manager = ActionManager([SHOOT, MOVE, DRIBBLE])
agent = QLearningAgent(num_states=features_manager.get_num_states(),
num_actions=actions_manager.get_num_actions(),
learning_rate=0.1,
discount_factor=0.99,
epsilon=1.0,
epsilon_dec=0.9992)
# Run training using Q-Learning
if train_mode == "train_only":
print('\n=== Train Mode for {}:'.format(num_train_ep))
train(num_episodes=num_train_ep,
game_interface=hfo_interface,
features=features_manager,
agent=agent,
actions=actions_manager)
elif train_mode == "alternate":
parser.add_argument('--saveFile', type=str, default="q_agent.model")
args = parser.parse_args()
num_teammates = args.numTeammates
num_opponents = args.numOpponents
n_games = args.numEpisodes
save_file = args.saveFile
lr = 0.01
epsilon = 1
discount_factor = 0.9
# Useful Instances:
hfo = HFOEnvironment()
hfo.connectToServer(feature_set=HIGH_LEVEL_FEATURE_SET, server_port=6000)
env = DiscreteHighLevelFeatures(hfo.getState(), num_teammates,
num_opponents)
actions = Action()
agent = QLearner(num_states=env.get_num_states(),
num_actions=actions.get_num_actions(),
epsilon=epsilon,
learning_rate=lr,
discount_factor=discount_factor,
save_file=save_file)
# Saving lists
scores = []
eps_history = []
for i in range(n_games):
print("\n<< {}/{} Game >> eps={}".format(i, n_games, agent.epsilon))
game_status = IN_GAME
score = 0
while game_status == IN_GAME:
action_idx = agent.choose_action(env.get_state_index())
hfo_action = actions.map_action(action_idx)
num_episodes, load_file))
# Initialize connection with the HFO server
hfo_interface = HFOAttackingPlayer(agent_id=agent_id,
num_opponents=args.num_opponents,
num_teammates=args.num_teammates)
hfo_interface.connect_to_server()
# Reward Function
reward_function = simple_reward
# Get number of features and actions
features_manager = DiscreteHighLevelFeatures(num_team, num_op)
actions_manager = ActionManager([SHOOT, MOVE, DRIBBLE])
# Initialize a Q-Learning Agent
agent = QPlayerAgent(num_states=features_manager.get_num_states(),
num_actions=actions_manager.get_num_actions(),
num_games=num_episodes,
load_file=load_file)
for i in range(num_episodes):
agent.reset(i, produce_graph=True)
observation = hfo_interface.reset()
# Update environment features:
curr_state_id = features_manager.get_state_index(observation)
has_ball = features_manager.has_ball(observation)
while hfo_interface.in_game():
action_idx = agent.act(curr_state_id)
hfo_action = actions_manager.map_action(action_idx)