agent = RDDPGAgent(state_size=state_size,
action_size=action_size,
actor_lr=args.actor_lr,
critic_lr=args.critic_lr,
tau=args.tau,
gamma=args.gamma,
lambd=args.lambd,
batch_size=args.batch_size,
memory_size=args.memory_size,
epsilon=args.epsilon,
epsilon_end=args.epsilon_end,
decay_step=args.decay_step,
load_model=args.load_model)
episode = 0
env = Env()
if args.play:
while True:
try:
done = False
bug = False
# stats
bestY, timestep, score, avgvel, avgQ = 0., 0, 0., 0., 0.
observe = env.reset()
image, vel = observe
try:
image = transform_input(image, args.img_height,
args.img_width)
targetY = 58
gamma = 0.99
lamb = 0.90
max_step = 600
score_bank = []
episode = 0
if __name__ == "__main__":
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
# Make RL agent
model = PPO_Network(SEQUENCE_SIZE, W, H, ACTION_SIZE)
functions = PPO_Functions()
env = Env()
#set session, and initialize
model.set_session(tf.Session(config=config))
try:
for e in range(5000000):
done = False
bug = False
level = -1
reward_sum = 0
t = 0
score = 0
observation = env.reset()
responses = observation[0]
quad_vel = observation[1]
# stack history here
try:
agent3 = RDQNAgent(
state_size=state_size,
action_size=action_size,
lr=args.lr,
gamma=args.gamma,
batch_size=args.batch_size,
memory_size=args.memory_size,
epsilon=args.epsilon,
epsilon_end=args.epsilon_end,
decay_step=args.decay_step,
load_model=args.load_model,
agent_name=agent_name3
)
episode = 0
env = Env()
if args.play:
while True:
try:
done = False
bug = False
# stats
bestReward, timestep, score, avgQ = 0., 0, 0., 0.
observe = env.reset()
image, vel = observe
vel = np.array(vel)
try:
image1 = transform_input(image[0], args.img_height, args.img_width)
quad_offset = (0, 0, -scaling_factor)
return quad_offset
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--verbose', action='store_true')
parser.add_argument('--continuous', action='store_true')
args = parser.parse_args()
if args.continuous:
agent = RandomAgentContinuous(3)
else:
agent = RandomAgentDiscrete(7)
env = Env()
episode = 0
while True:
done = False
timestep = 0
score = 0
_ = env.reset()
while not done:
timestep += 1
action = agent.get_action()
if not args.continuous:
action = interpret_action(action)
_, reward, done, info = env.step(action)
score += reward
