#Catch KeyboardInterrupts and save model
#i=-1
# try:
# #Reinforcement Loop
# #for i in tqdm.trange(n_episodes):
# while True:
# i += 1
for i in range(n_episodes):
info, reward, state = env.reset(
) # reset env before starting a new episode
j = 0
shortterm_memory = ReplayMemory(max_size=256)
while True:
j += 1
# interact with env
action = agent.step(state)
#observation, reward, done, info = env.step(action)
done, base_reward, observation = env.step(action)
#Determine real reward based on Policy
reward = Policies.LiveLongAndProsper(base_reward, done)
#reward = base_reward
# store transaction in memory
transition = [state, action, reward, observation, done]
shortterm_memory.store(*transition)
memory.store(*transition)
# Step to next state
state = observation
if torch.cuda.is_available():
torch.cuda.empty_cache()
print("Running on GPU")
agent.target_net.cuda()
agent.policy_net.cuda()
#Reinforcement Loop
for i in range(n_episodes):
info, reward, state = env.reset(
) # reset env before starting a new episode
j = 0
while True:
j += 1
# interact with env
action = agent.step(state, decay_enabled=False)
#observation, reward, done, info = env.step(action)
done, reward, observation = env.step(action)
#Determine real reward based on Policy
#reward = Policies.SoreLoser(reward, done)
# Step to next state
state = observation
#Save rewards for evaluation
R[i] = reward
#Reset if game lasts too long:
#Protects against environment bug where agents can be trapped outside the arena