def train(self):
self.reset()
score_logger = ScoreLogger(self.env_name)
for run in range(self.max_iter):
state = self.env.reset(self.num)
if self.log_state:
print("STATE")
print(state)
state = np.reshape(state, [1, self.observation_space])
step = 0
while self.num < self.max_examples:
step += 1
#self.env.render()
action = self.act(state)
state_next, reward, terminal, info = self.env.step(action, self.num)
if self.log_state:
print(action, reward, terminal)
#reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, self.observation_space])
self.remember(state, action, reward, state_next, terminal)
if self.num >= self.max_examples:
terminal = True
state = state_next
if terminal:
str = "Run: {:d}, exploration: {:.5f}, score: {:d}, memory: {:d}, balance: {:.2f}"
print(str.format(run+1, self.exploration_rate, step, len(self.memory), self.env.getBalance()))
score_logger.add_score(step, run+1)
break
self.experience_replay()
if self.num >= self.max_examples:
break
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space]) # reshape to row array
step = 0
while True: # run the Deep-Q Net
step += 1
env.render()
action = dqn_solver.act(state) # make action according to e-greedy
state_next, reward, terminal, info = env.step(action) # observe SARS from env
reward = reward if not terminal else -reward # ? negative
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal) # store transition
state = state_next # advance to next state
if terminal:
print("Run: " + str(run) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(step))
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
if run % SAVE_FREQUENCY == 0:
save_model(dqn_solver.model, "saved_model") # save the model parameters every once in a while
def cartpole(iteration=0, params=None):
if params is None:
params = DefaultParams()
env = gym.make(params.ENV_NAME)
score_logger = ScoreLogger(params.ENV_NAME, iteration, params.EXP_NAME,
params.FIXED_NB_RUNS)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space, params)
run = 0
done = False
while done == False:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
#env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step))
done = score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
def cartpole():
env = gym.make(ENV_NAME)
env.spec.reward_threshold = 1000
env.spec.max_episode_steps = 1000
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
if LOAD and os.path.exists(LOAD_MODULE):
dqn_solver.load_mode()
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print(
"Run: " + str(run) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(step))
if len(score_logger.scores) > 0 and step > max(score_logger.scores) > 450:
dqn_solver.save_model(step)
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
def cartpole():
env = Env('localhost:32822')
env.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.box.shape[0]
action_space = env.action_space.discrete.n
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
while True:
run += 1
state = env.reset()
# print(state)
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
# env.render()
print("acting on state: ", state)
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(step))
score_logger.add_score(step, run)
plt.plot(dqn_solver.loss)
plt.title('Model loss')
plt.ylabel('Loss')
plt.xlabel('Episode')
plt.savefig("loss.png")
break
dqn_solver.experience_replay()
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(action_space, is_partial_fit=False)
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
# comment next line for faster learning, without stopping to show the GUI
env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(step))
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
# NEEDED TO RUN CODE ONLY ON CPU ON DGX
config = tf.ConfigProto(
device_count = {'GPU': 1}
)
sess = tf.Session(config=config)
set_session(sess)
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
#env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(step))
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
#dqn_solver.load("./save/cartpole-dqn.h5")
run = 0
for e in range(EPISODES):
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
#env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(step))
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
if step % 5 == 0:
dqn_solver.save("./save/cartpole-dqn_step_{}.h5".format(step))
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space, False)
run = 0
# while True:
for i in range(200):
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
print(state_next)
state_next = np.reshape(state_next, [1, observation_space])
print(state_next)
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step))
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
dqn_solver.save_model_and_table()
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
#env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step))
score_logger.add_score(step, run)
break
if step % TRAINING_STEP == 1:
dqn_solver.experience_replay()
def mountain_car():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(action_space, is_partial_fit=True)
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
max_position = -1.2
while True:
step += 1
# comment next line for faster learning, without stopping to show the GUI
#env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
# if max_position < state_next[0]:
# reward = 0
# if terminal and reward > 0:
# reward = 100
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
max_position = max(max_position, state[0][0])
if terminal:
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step) + ", max: " + str(max_position))
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
def cartpole2():
env = gym.make(ENV_NAME)
env._max_episode_steps = 700
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
score = [[0, 0, 0]]
agents_weights = []
check = [50, 100, 200, 400, 500, 1000, 2000, 3000]
while mean_score_check(np.mean(score, axis=1)) < 500:
run += 1
score_times = []
for i in range(n_times):
step = 0
state = env.reset()
state = np.reshape(state, [1, observation_space])
done = False
if i == 0:
while (not done):
step += 1
#env.render()
action = dqn_solver.act(state)
agents_weights.append(dqn_solver.model_q.get_weights())
state_next, reward, done, info = env.step(action)
reward = reward if not done else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next,
done)
state = state_next
dqn_solver.experience_replay()
else:
while (not done):
step += 1
# env.render()
action = dqn_solver.act_eval(state)
#agents_weights.append(dqn_solver.model_q.get_weights())
state_next, reward, done, info = env.step(action)
reward = reward if not done else -reward
state_next = np.reshape(state_next, [1, observation_space])
state = state_next
if done:
score_times.append(step)
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(np.mean(score_times)))
score_logger.add_score(int(np.mean(score_times)), run)
score.append(score_times)
#print(score_times)
if run in check:
np.save('weights', np.array(agents_weights))
np.save('scores', np.array(score))
np.save('weights', np.array(agents_weights))
np.save('scores', np.array(score))
def cartpole():
#Creating the gym environment
env = gym.make(ENV_NAME)
env = gym.wrappers.Monitor(env, "./vid", video_callable=lambda episode_id: True, force=True)
#initializing the score logger to visualize later on the performance of the AI
score_logger = ScoreLogger(ENV_NAME)
"""
Observation:
Type: Box(4)
Num Observation Min Max
0 Cart Position -4.8 4.8
1 Cart Velocity -Inf Inf
2 Pole Angle -24 deg 24 deg
3 Pole Velocity At Tip -Inf Inf
Actions:
Type: Discrete(2)
Num Action
0 Push cart to the left
1 Push cart to the right
"""
#There are 4 possible observations in this environment, so observation space will be equal to 4
observation_space = env.observation_space.shape[0]
#There are 2 possible actions, moving to the left and moving to the right as seen above
action_space = env.action_space.shape[0]
#DQNSolver is the "AI", the agent that from the list of observations and actions will try to determine the best actions for given circumstances (observation)
#Initializing the dqn_solver
dqn_solver = DQNSolver(observation_space, action_space)
#Run is a variable to track how many runs it has been
run = 0
while True:
run += 1
#When you call env.reset(), it returns the initial state as a np.ndarray of shape (4,) since there are 4 observations
state = env.reset()
#Reshaping the state into a 2d array of (1,4)
state = state.reshape(1, observation_space)
step = 0
while True:
#Each step is a new action undertaken by the agent
step += 1
env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(step))
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
env.close()
def cartpole(GAMMA, LEARNING_RATE, EXPLORATION_DECAY):
#environment - cartpole
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
#observational space - possible state values
observation_space = env.observation_space.shape[0]
#action space - possible actions that can be performed
action_space = env.action_space.n
#agent - object of DQN Solver class, see below
dqn_solver = DQNSolver(observation_space, action_space, GAMMA, LEARNING_RATE, EXPLORATION_DECAY)
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
#visualize environment
#env.render()
#determine action
action = dqn_solver.act(state)
#determine new state and corresponding reward
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
#remember to learn - used in experience replay
dqn_solver.remember(state, action, reward, state_next, terminal)
#set future state as current state
state = state_next
if terminal:
#Add score to score logger once pendulum falls
score_logger.add_score(step, run)
break
#calling experience replay to update Q value
dqn_solver.experience_replay(GAMMA, EXPLORATION_DECAY)
#return runs to solve
if score_logger.solve_score != 0:
return score_logger.solve_score
def cartpole():
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
port = 12345
# connection to hostname on the port.
s.connect(('192.168.43.110', port))
print("Connected to environment")
score_logger = ScoreLogger(ENV_NAME)
observation_space = 4
action_space = 4
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
for _ in range(10):
run += 1
func = dict()
func["function"] = "render"
s.send(pickle.dumps(func))
state = pickle.loads(s.recv(1024))["state"]
step = 0
while True:
startTime = time.time()
step += 1
state = np.reshape(state, [1, observation_space])
action = dqn_solver.act(state)
# action = 0
func["function"] = "step"
func["action"] = action
s.send(pickle.dumps(func))
recieved = pickle.loads(s.recv(1024))
state_next, reward, terminal = recieved["state"], recieved[
"reward"], recieved["terminal"]
print(state, reward, terminal)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
print(state_next.shape)
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal or time.time() - startTime > 30:
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step))
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
dqn_solver.save_model_and_table()
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(action_space, is_partial_fit=True)
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
# run = 0
while True:
step += 1
env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(
state_next,
[1, observation_space
]) # Gives a new shape to an array without changing its data.
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration_rate: " +
str(dqn_solver.exploration_rate) + ", episodes: " +
str(step))
# score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
def mountaincar():
# initialize game and score logger
env = gym.make(ENV_NAME)
# tool created to display 'score'
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
score = 0
numTries = 200
# continue indefinitely to train and perfect the mdoel
while run < numTries:
run += 1
# a new frame
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
# keep stepping til you fall out of frame
while True:
# more steps = more successful (even if it is moving)
env.render()
# choose what to do using DQN
action = dqn_solver.act(state)
# analyze what happened
state_next, reward, dead, info = env.step(action)
# reinforce positive outcome, penalize bad outcome
reward = get_reward(state_next)
step += reward
state_next = np.reshape(state_next, [1, observation_space])
# memorize this iteration to shape the following
dqn_solver.memorize(state, action, reward, state_next, dead)
state = state_next
if dead:
# score = # of steps taken in a particular run (too many steps is bad)
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step))
score_logger.add_score(step, run)
break
dqn_solver.replay()
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
# Size of observations
observation_space = env.observation_space.shape[0]
# Size of actions
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
game_nb = 0
while True:
game_nb += 1
# Game creation/reset
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
finished = False
# While game does not end
while not finished:
step += 1
env.render()
# Decide which move
action = dqn_solver.act(state)
# Perform the move
state_next, reward, finished, info = env.step(action)
state_next = np.reshape(state_next, [1, observation_space])
# If the game has finished (we failed) invert the reward
reward = reward if not finished else -reward
# Save the movement and its reward
dqn_solver.remember(state, action, reward, state_next, finished)
state = state_next
# Learn from experience
dqn_solver.experience_replay()
if finished:
# Save result
print ("Game number: " + str(game_nb) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(step))
score_logger.add_score(step, game_nb)
def training():
env = gym.make(ENV_NAME)
# If the user chooses an environment with a non-discrete action space, return an error because DQN only works with discrete action spaces
if (type(env.action_space) != gym.spaces.discrete.Discrete):
raise ActionSpaceError(
'This environment uses an action space that is not discrete. DQN can only be trained using discrete action spaces. Please select an envionment with a discrete action space.'
)
act_space = env.action_space.n
score_logger = ScoreLogger(ENV_NAME)
observation_input = find_input_shape(env)
dims = reshape_dims(env.observation_space)
dqn_solver = DQNSolver(observation_input, act_space)
for i in range(NUM_EPISODES):
state = env.reset()
#reshape state array if it has more than one dimension
if (len(dims) > 1):
state = state.reshape(dims)
step = 0
while True:
step += 1
if (WATCH_TRAINING):
env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
#reshape state array if it has more than one dimension
if (len(dims) > 1):
state_next = state_next.reshape(dims)
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
if terminal:
print("Run: " + str(i + 1) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step))
score_logger.add_score(step, i + 1)
break
dqn_solver.experience_replay()
return dqn_solver
def main():
# Kornyezet import
env = gym.make(ENV_NAME)
# Pont szamolas
score_logger = ScoreLogger(ENV_NAME)
# Observation space + action space definialas
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
# RL modell
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
dqn_solver.add_to_memory(state, action, reward, state_next,
terminal)
state = state_next
if terminal:
result = "Run: " + str(run) + ", exploration: " + str(
dqn_solver.exploration_rate) + ", score: " + str(step)
print(result)
score_logger.add_score(step, run)
break
dqn_solver.experience()
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(action_space, is_partial_fit=IS_PARTIAL_FIT)
run = 0
current_max = 0
while run < NO_ITERATIONS:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
# run = 0
while True:
step += 1
if RENDER:
env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
if MY_LOGS:
print(state_next, reward, terminal)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next,
[1, observation_space]) # Gives a new shape to an array without changing its data.
dqn_solver.remember(state, action, reward, state_next, terminal)
state = state_next
current_max = max(current_max, step)
if terminal:
print("Run: " + str(run) + ", exploration_rate: " + str(
dqn_solver.exploration_rate) + ", episodes: " + str(step), ", max episodes reached: ",
str(current_max))
if (SCORE_LOGGER):
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
def main():
n_episode = 50
for method in ['DoubleDQN']:
env = gym.make(ENV_NAME)
agent = None
if method == 'DDPG':
agent = DDPG(env)
elif method == 'QLeaning':
agent = QLearningTabular(env)
elif method == 'DQN':
agent = DQN(env)
elif method == 'DoubleDQN':
agent = DoubleDQN(env)
else:
raise NotImplementedError
score_logger = ScoreLogger(ENV_NAME, method)
print('Algorithm:', method)
i_episode = 0
while i_episode < n_episode:
i_episode += 1
agent.run_begin()
state = env.reset()
i_step = 0
while True:
i_step += 1
# env.render()
action = agent.act(state)
state_next, reward, terminal, info = env.step(action)
agent.remember(state, action, reward, state_next, terminal)
agent.experience_replay()
state = state_next
if terminal:
agent.run_finish()
print(f"Run {i_episode} finished after {i_step + 1} steps")
score_logger.add_score(i_step, i_episode)
break
env.close()
def cartpole():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space
action_space = env.action_space
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space.shape[0]])
step = 0
over2 = False
over0 = False
over3 = False
while True:
step += 1
action = dqn_solver.act(state)
env.render()
state_next, reward, terminal, info = env.step(action)
# reward = reward if not terminal else -20
if state_next[0] > -0.2 and not over2:
over2 = True
reward = 1
if state_next[0] > 0 and not over0:
over0 = True
reward = 1
if state_next[0] > 0.2 and not over3:
over3 = True
reward = 1
if state_next[0] >= 0.5:
reward = 20
# reward = -reward
state_next = np.reshape(state_next,
[1, observation_space.shape[0]])
dqn_solver.remember(state, action, reward, state_next, terminal)
dqn_solver.experience_replay()
dqn_solver.target_train()
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(reward))
print("state: " + str(state_next[0, 0]))
# score_logger.add_score(reward, run)
break
def cartpole():
env = gym.make("CartPole-v1")
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
dqn_solver.build_fresh()
# dqn_solver.load_model()
run = 0
while True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
step += 1
env.render()
action = dqn_solver.act(state)
state_next, reward, terminal, info = env.step(action)
reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
if dqn_solver.train_mode:
dqn_solver.remember(state, action, reward, state_next,
terminal)
state = state_next
if terminal:
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step))
is_solved = score_logger.add_score(step, run)
if is_solved and dqn_solver.train_mode:
dqn_solver.save_model()
exit()
break
if dqn_solver.train_mode:
dqn_solver.experience_replay()
def traffic():
env = gym.make(ENV_NAME)
#env = gym.wrappers.Monitor(env, "dqn")
env.seed(1)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
while run < 1:
run += 1
step = 0
state = env._reset()
obs0, reward_previous, don, signal = state
reward_current = 0
total_reward = reward_previous - reward_current
print("STATUS")
print(signal)
if (signal == 0):
status = 0
elif (signal == 1):
status = 1
next_state = obs0
while step < 1000:
#env.render()
step += 1
action = dqn_solver.act(state)
#print(next_state)
#action = env.action_space.sample()
#obs1, reward_previous, done, _ = env.step(action)
if (status == 0 and action == 0):
print("Status is: 0. Action is 0.")
status = 0
next_state, reward_current, done, _, t_step = phase(env, 0, 15)
step += t_step
elif (status == 0 and action == 1):
print("Status is 0. Action is now 1. Switching to Status 1.")
phase(env, 2, 25)
#print("Action is 1. Status is 0. Lights are H-Y, V-R -> H-R, V-G")
status = 1
next_state, reward_current, done, _, t_step = phase(env, 1, 45)
step += t_step
elif (status == 1 and action == 1):
print("Status is 1. Action is 1.")
status = 1
next_state, reward_current, done, _, t_step = phase(env, 1, 15)
step += t_step
elif (status == 1 and action == 0):
print("Status is 1. Action is now 0. Switching to Status 0.")
phase(env, 4, 25)
status = 0
next_state, reward_current, done, _, t_step = phase(env, 0, 45)
step += t_step
total_reward = reward_previous - reward_current
state_next = np.reshape(next_state, [1, OBS_SPACE])
dqn_solver.remember(state, action, total_reward, state_next, done)
state = state_next
score_logger.add_score(step, run)
print(step)
if done:
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step))
#score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
print("Episode done in %d steps, total reward %.2f" % (step, total_reward))
env.close()
def cartpole():
#Creating the gym environment
env = gym.make(ENV_NAME)
#saving all the episodes to a folder vid in a video format
recorder = gym.wrappers.monitoring.video_recorder.VideoRecorder(env, base_path="./every_100")
#initializing the score logger to visualize later on the performance of the AI
score_logger = ScoreLogger(ENV_NAME)
"""
Observation:
Type: Box(4)
Num Observation Min Max
0 Cart Position -4.8 4.8
1 Cart Velocity -Inf Inf
2 Pole Angle -24 deg 24 deg
3 Pole Velocity At Tip -Inf Inf
Actions:
Type: Discrete(2)
Num Action
0 Push cart to the left
1 Push cart to the right
"""
#There are 4 possible observations in this environment, so observation space will be equal to 4. These are game specific inputs
observation_space = env.observation_space.shape[0]
#There are 2 possible actions, moving to the left and moving to the right as seen above
action_space = env.action_space.n
#DQNSolver is the "AI", the agent that from the list of observations and actions will try to determine the best actions for given circumstances (observation)
#Initializing the dqn_solver
dqn_solver = DQNSolver(observation_space, action_space)
#Run is a variable to track how many runs it has been
run = 0
while True:
run += 1
#When you call env.reset(), it returns the initial state as a np.ndarray of shape (4,) since there are 4 observations
state = env.reset()
#Reshaping the state into a 2d array of (1,4)
state = state.reshape(1, observation_space)
step = 0
while True:
#Each step is a new action undertaken by the agent
step += 1
#Comment this line below if you don't want to see the animation rendered live. It will make it the computations quicker
if run % RECORD_EVERY == 0:
# env.render()
recorder.capture_frame() #Return an action based on the state
action = dqn_solver.act(state)
#Environment returns informations based on the action decided by the dqn_solver
state_next, reward, terminal, info = env.step(action)
#Positive or negative reward
reward = reward if not terminal else -reward
#The next state
state_next = state_next.reshape(1, observation_space)
#This is how our AI has "memory"
dqn_solver.remember(state, action, reward, state_next, terminal)
#update the state from the past to present
state = state_next
#Terminal means the episode/run has finished
if terminal:
print("Run: " + str(run) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(step))
score_logger.add_score(step, run)
break
dqn_solver.experience_replay()
env.close()
def runCat3D():
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
'''
method = 1: use DQNSolver
method = 2: use cyclic_action, which makes the cat land on its feet (under the current initial conditions and parameters)
method = 3: user inputs the actions
'''
method = 1
if method == 1:
print("Using DQNSolver.")
dqn_solver = DQNSolver(observation_space, action_space)
fileWeights = "weights3D_simpler.h5"
#uncomment to start off with saved weights
#dqn_solver.load_weights(fileWeights)
elif method == 2:
print("Using cyclic method.")
else:
print("User inputs.")
run = 0
average_score = 0
while run < 100: #True:
run += 1
state = env.reset()
state = np.reshape(state, [1, observation_space])
step = 0
while True:
env.render()
if method == 1:
action = dqn_solver.act(state)
elif method == 2:
action = cyclic_action(step, env)
else:
action = input("Enter action")
action = int(action)
state_next, reward, terminal, info = env.step(action)
reward = -reward
#reward = reward if not terminal else -reward
state_next = np.reshape(state_next, [1, observation_space])
if method == 1:
dqn_solver.remember(state, action, reward, state_next,
terminal)
state = state_next
if terminal:
average_score += reward
if method == 1:
print("Run: " + str(run) + ", exploration: " +
str(round(dqn_solver.exploration_rate, 4)) +
", score: " + str(round(reward, 2)))
else:
print("Run: " + str(run) + ", score: " +
str(round(reward, 2)))
#score_logger.add_score(int(reward), run)
break
if method == 1:
dqn_solver.experience_replay()
if run % 50 == 0:
dqn_solver.save_weights(fileWeights)
step += 1
print("Total runs: " + str(run) + ", average score: " +
str(round(average_score / run, 2)))
input("End. Press any key")
env.close()
def connect4dqn():
env = Connect4()
score_logger = ScoreLogger('Connect4')
player1won = 0
player2won = 0
observation_space = env.reset().shape
action_space = env.validMoves().size
# Assign GPU to DGX
config = tf.ConfigProto(device_count={'GPU': 2})
sess = tf.Session(config=config)
set_session(sess)
dqn_solver = DQNSolver(observation_space, action_space)
run = 0
state = env.reset(
) #moved one loop up. otherwise player two wont be able to start if player one wins
while True:
run += 1
if run % 50 == 0:
print('Saving weights and starting evaluation...')
dqn_solver.save()
score, ties = evaluate_dqn(env, dqn_solver, 1000)
score_logger.add_score(score + ties, run) #logging ties as success
step = 0
while True:
step += 1
player = env.getNextPlayer()
if player == 1:
action_player1 = dqn_solver.act(state, env)
state_next, reward_player1, terminal, info = env.makeMove(
player, action_player1)
state_copy = np.copy(state)
state_next_copy = np.copy(state_next)
if terminal:
dqn_solver.pop(
) # if player 1 wins, pop player 2's last move from and give it a negative reward
dqn_solver.remember(normalized_state, action_player2,
reward_player1 * -1,
normalized_state_next, terminal)
dqn_solver.remember(state, action_player1, reward_player1,
state_next, terminal)
state = state_next
else:
normalized_state = np.roll(state, 1, axis=-1)
action_player2 = dqn_solver.act(normalized_state, env)
# userInput = int(input("Which row silly Human? "))
# action_player2 = userInput
state_next, reward_player2, terminal, info = env.makeMove(
player, action_player2)
normalized_state_next = np.roll(state_next, 1, axis=-1)
if terminal:
dqn_solver.pop(
) # if player 2 wins, pop player 1's last move from and give it a negative reward
dqn_solver.remember(state_copy, action_player1,
reward_player2 * -1, state_next_copy,
terminal)
dqn_solver.remember(normalized_state, action_player2,
reward_player2, normalized_state_next,
terminal)
state = state_next
if terminal:
if player == 1:
player1won += 1
else:
player2won += 1
try:
winRatio = player1won / player2won
except ZeroDivisionError:
winRatio = 0
print('Win ratio: {}'.format(winRatio))
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", moves: " +
str(step))
break
for i in range(20):
dqn_solver.experience_replay()
def cartpole():
env = UnityEnv(environment_filename=ENV_NAME,
worker_id=2,
use_visual=False,
multiagent=True)
score_logger = ScoreLogger(ENV_NAME)
agents_brain = []
agents_action = []
index_list = []
agents_alive = []
count = 0
count1 = 0
num_agents = env.number_agents
print("___________Number of agents in cartpole __")
print(num_agents)
observation_space = env.observation_space.shape[0]
action_space = env.action_space.n
dqn_solver = DQNSolver(observation_space, action_space)
print("__dqn solver______")
print(dqn_solver)
#model = tf.keras.models.load_model("")
for x in range((env.number_agents)):
agents_brain.append(dqn_solver)
print("______agentbrain____")
print(agents_brain)
print("_Agent action___")
print(agents_action)
learning_brain = copy.deepcopy(agents_brain)
run = 0
state = env.reset()
initialstate = copy.deepcopy(state)
while True:
run += 1
env.reset()
print("____________STATE____________-")
print(state[0])
state = copy.deepcopy(initialstate)
agents_brain = []
agents_action = []
index_list = []
agents_alive = []
count = 0
count1 = 0
num_agents = int(state[0][-5])
agents_brain = copy.deepcopy(learning_brain)
print(learning_brain)
print(agents_brain)
print(state)
#for x in range ( (env.number_agents - 1) ):
step = 0
while True:
step += 1
env.render()
print("___________STatte Lenth_______")
print(len(state))
print("______selffish___")
print(state[0])
agents_action = [1] * len(state)
copied_agents_alive = copy.deepcopy(agents_alive)
print("__________numagents_____")
for x in range(num_agents - 1):
state[x] = np.reshape(state[x], [1, observation_space])
agents_action[x] = agents_brain[x].act(state[x])
print(agents_action)
state_next, reward, terminal, info = env.step(
agents_action, num_agents)
print("_______Reward________")
print(reward)
print("_____________NEXT STATE LENGTH____________")
print(len(state_next))
if (len(state_next) == 0):
break
agents_alive = state_next[0][-13:-5]
num_agents = int(state_next[0][-5])
print("_______num agnets in cartpole________")
print(num_agents)
print("_____index list")
print(index_list)
print(agents_alive)
agents_alive1 = np.delete(agents_alive, index_list)
print("_______Alive agent list_______")
print(agents_alive1)
flag = False
# del agents_alive[index_list[x]]
for x in range(len(agents_alive)):
if (agents_alive[x] == float(1)):
for y in range(len(index_list)):
if (index_list[y] == x):
flag = True
if (flag == False):
index_list.append(x)
flag = False
index_to_remove = []
for x in range(len(agents_alive1)):
if (agents_alive1[x] == float(1)):
learning_brain[index_list[count]] = agents_brain[x]
index_to_remove.append(x)
count = count + 1
agents_brain = [
i for j, i in enumerate(agents_brain)
if j not in index_to_remove
]
print("____________AGENTS_BRAIN_________")
print(len(agents_brain))
print("_______________Terminal_____________")
print(terminal)
if (terminal[0] == True):
print("Run: " + str(run) + ", exploration: " +
str(dqn_solver.exploration_rate) + ", score: " +
str(step))
score_logger.add_score(step, run)
for x in range(len(copied_agents_alive)):
learning_brain[x] = agents_brain[count1]
count1 = count1 + 1
for x in range(len(learning_brain)):
learning_brain[x].save(str(run) + "brain" + str(x) + ".h5")
break
for x in range(num_agents - 1):
state[x] = np.reshape(state[x], [1, observation_space])
state_next[x] = np.reshape(state_next[x],
[1, observation_space])
agents_brain[x].remember(state[x], agents_action[x], reward[x],
state_next[x], terminal[x])
agents_brain[x].experience_replay()
state = state_next
MEMORY_SIZE = 1000000
BATCH_SIZE = 32
EXPLORATION_MAX = 1.0
EXPLORATION_MIN = 0.01 # 1% of the time the agent will explore
EXPLORATION_DECAY = 0.995
N_EPISODES = 1001
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1, 2"
# Set Parameters
env = gym.make(ENV_NAME)
score_logger = ScoreLogger(ENV_NAME)
state_size = env.observation_space.shape[0]
action_size = env.action_space.n
output_dir = 'model_output/cartpole'
if not os.path.exists(output_dir):
os.makedirs(output_dir)
class DQNNN(nn.Module):
def __init__(self, state_size, action_size):
super(DQN, self).__init__()
self.dense1 = nn.Linear(state_size, 24)
self.dense2 = nn.Linear(24, 24)
self.output = nn.Linear(23, action_size)