def test_transition2():
initial_state = SoccerEnv.encode_state(0,3,0,1,False)
action = SoccerEnv.encode_action(SoccerEnv.Action.S, SoccerEnv.Action.N)
env = SoccerEnv()
transitions = env.P[initial_state][action]
expected_next_state = SoccerEnv.encode_state(1,3,0,1,False)
for prob, next_state, reward, done in transitions:
assert next_state == expected_next_state
assert reward == 0
assert done == 0
def test_transitions():
transitions = SoccerEnv.transitions(0, 2, 0, 1, 0, SoccerEnv.Action.W, SoccerEnv.Action.Stick)
expected_states = set([SoccerEnv.encode_state(0, 2, 0, 1, 0), SoccerEnv.encode_state(0, 2, 0, 1, 1)])
assert len(transitions) == 2
for next_state, reward, done in transitions:
assert next_state in expected_states
assert reward == 0
assert done == 0
state = SoccerEnv.encode_state(0, 2, 0, 1, 0)
action = SoccerEnv.encode_action(SoccerEnv.Action.W, SoccerEnv.Action.Stick)
env = SoccerEnv()
transitions = env.P[state][action]
assert len(transitions) == 2
for prob, next_state, reward, done in transitions:
assert abs(prob - 0.5) < 0.001
assert next_state in expected_states
assert reward == 0
assert done == 0
n_episodes_MAX = 2 * 10**5 #10*10**5
steps_MAX = 100
verbose = False
alpha = 0.1
# Instantiate Q-Learner
from QLearner import QLearningAgent
Foe_Q_agent = QLearningAgent(state_size=num_states,
action_size=num_individual_actions,
value_state_function_learner='minimax-q',
learning_rate=alpha)
# Ref state s, action 'S'
ref_state = SoccerEnv.encode_state(0, 2, 0, 1, 0)
ref_P1_action = int(SoccerEnv.Action.S)
ref_P2_action = int(SoccerEnv.Action.Stick)
# Q errors for plotting
Foe_Q_P1_Q_errors = []
for i_episode in range(n_episodes_MAX):
state = env.reset()
P1_Q_ref = Foe_Q_agent.Q[ref_state, ref_P1_action, ref_P2_action]
for t in range(steps_MAX):
joint_action = np.random.randint(num_actions)
# Take action A, observe R, S'