def test_render():
env = SoccerEnv()
env.render()
action = env.encode_action(SoccerEnv.Action.Stick, SoccerEnv.Action.Stick)
env.step(action)
env.render()
return
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
def test_action_encode():
env = SoccerEnv()
action1, action2 = 1, 2
x = env.encode_action(1,2)
assert (action1, action2) == env.decode_action(x)
ref_state = SoccerEnv.encode_state(0, 2, 0, 1, 0)
ref_P1_action = int(SoccerEnv.Action.S)
# Q errors for plotting
Q_Learner_P1_Q_errors = []
for i_episode in range(n_episodes_MAX):
state = env.reset()
P1_Q_ref = Q_agent.Q[ref_state, ref_P1_action]
for t in range(steps_MAX):
P1_action = Q_agent.sample_policy(state)
P2_action = np.random.randint(num_actions)
joint_action = env.encode_action(P1_action, P2_action)
# Take action A, observe R, S'
state_new, reward, done, info = env.step(joint_action)
Q_agent.learn(reward, state, state_new, P1_action)
state = state_new
if done:
if verbose:
print("Episode finished after {} timesteps".format(t + 1))
break
# calc error at end of episode update
Q_Learner_P1_Q_errors.append(np.abs(Q_agent.Q[ref_state, ref_P1_action] - P1_Q_ref))