def __init__(self): self.BATCH_SIZE = 128 self.GAMMA = 0.99 self.EPS_START = 1.0 self.EPS_END = 0.05 self.EPS_DECAY = 0.000005 self.TARGET_UPDATE = 5 self.pretrain_length = self.BATCH_SIZE #self.state_size = [55,3] self.action_size = 3 self.hot_actions = np.array(np.identity(self.action_size).tolist()) #self.action_size = len(self.hot_actions) self.learning_rate = 0.0005 #self.total_episodes = 12 self.max_steps = 1000 self.env = Environment() self.memory_maxsize = 10000 self.DQNetwork = DQNetwork(learning_rate = self.learning_rate,name = 'DQNetwork') self.TargetNetwork = DQNetwork(learning_rate = self.learning_rate , name = 'TargetNetwork') self.memory = ReplayMemory(max_size=self.memory_maxsize) self.saver = tf.train.Saver()
def __init__(self,
state_size,
action_size,
seed=0,
buffer_size=100000,
batch_size=64,
update_frequency=2,
gamma=.99,
learning_rate=5e-4,
tau=1e-3):
self.state_size = state_size
self.action_size = action_size
self.random = random.seed(seed)
self.batch_size = batch_size
self.memory = ReplayBuffer(self.action_size, buffer_size, batch_size,
seed)
self.time_step = 0
self.update_frequency = update_frequency
self.qnetwork_local = DQNetwork(state_size, action_size,
seed).to(device)
self.qnetwork_target = DQNetwork(state_size, action_size,
seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(),
lr=learning_rate)
# hyper-parameters
self.gamma = gamma
self.tau = tau
def __init__(self,
lr,
inputChannels,
stateShape,
numActions,
batchSize,
epsilon=1.0,
gamma=0.99,
layer1Size=1024,
layer2Size=512,
maxMemSize=100000,
epsMin=0.01,
epsDecay=5e-4):
self.lr = lr
self.epsilon = epsilon
self.epsMin = epsMin
self.epsDecay = epsDecay
self.gamma = gamma
self.batchSize = batchSize
self.actionSpace = list(range(numActions))
self.maxMemSize = maxMemSize
self.memory = ReplayBuffer(maxMemSize, stateShape)
self.deepQNetwork = DQNetwork(lr, inputChannels, numActions)
def __init__(self, state_size, action_size, seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = DQNetwork(state_size, action_size,
seed).to(device)
self.qnetwork_target = DQNetwork(state_size, action_size,
seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
class DQAgent():
def __init__(self,
lr,
inputChannels,
stateShape,
numActions,
batchSize,
epsilon=1.0,
gamma=0.99,
layer1Size=1024,
layer2Size=512,
maxMemSize=100000,
epsMin=0.01,
epsDecay=5e-4):
self.lr = lr
self.epsilon = epsilon
self.epsMin = epsMin
self.epsDecay = epsDecay
self.gamma = gamma
self.batchSize = batchSize
self.actionSpace = list(range(numActions))
self.maxMemSize = maxMemSize
self.memory = ReplayBuffer(maxMemSize, stateShape)
self.deepQNetwork = DQNetwork(lr, inputChannels, numActions)
'''
REENABLE EPSILON GREEDY
'''
def chooseAction(self, observation):
if np.random.random() > self.epsilon:
state = torch.tensor(observation).float().clone().detach()
state = state.to(self.deepQNetwork.device)
state = state.unsqueeze(0)
policy = self.deepQNetwork(state)
action = torch.argmax(policy).item()
return action
else:
return np.random.choice(self.actionSpace)
def storeMemory(self, state, action, reward, nextState, done):
self.memory.storeMemory(state, action, reward, nextState, done)
def learn(self):
if self.memory.memCount < self.batchSize:
return
self.deepQNetwork.optimizer.zero_grad()
stateBatch, actionBatch, rewardBatch, nextStateBatch, doneBatch = \
self.memory.sample(self.batchSize)
stateBatch = torch.tensor(stateBatch).to(self.deepQNetwork.device)
actionBatch = torch.tensor(actionBatch).to(self.deepQNetwork.device)
rewardBatch = torch.tensor(rewardBatch).to(self.deepQNetwork.device)
nextStateBatch = torch.tensor(nextStateBatch).to(
self.deepQNetwork.device)
doneBatch = torch.tensor(doneBatch).to(self.deepQNetwork.device)
batchIndex = np.arange(self.batchSize, dtype=np.int64)
actionQs = self.deepQNetwork(stateBatch)[batchIndex, actionBatch]
allNextActionQs = self.deepQNetwork(nextStateBatch)
nextActionQs = torch.max(allNextActionQs, dim=1)[0]
nextActionQs[doneBatch] = 0.0
qTarget = rewardBatch + self.gamma * nextActionQs
loss = self.deepQNetwork.loss(qTarget,
actionQs).to(self.deepQNetwork.device)
loss.backward()
self.deepQNetwork.optimizer.step()
if self.epsilon > self.epsMin:
self.epsilon -= self.epsDecay
def __init__(self, level_name):
self.level_name = level_name
# setup environment
self.env = gym_super_mario_bros.make(level_name)
self.env = JoypadSpace(self.env, SIMPLE_MOVEMENT)
# one hot encoded version of our actions
self.possible_actions = np.array(np.identity(self.env.action_space.n, dtype=int).tolist())
# resest graph
tf.reset_default_graph()
# instantiate the DQNetwork
self.DQNetwork = DQNetwork(state_size, action_size, learning_rate)
# instantiate memory
self.memory = Memory(max_size=memory_size)
# initialize deque with zero images
self.stacked_frames = deque([np.zeros((100, 128), dtype=np.int) for i in range(stack_size)], maxlen=4)
for i in range(pretrain_length):
# If it's the first step
if i == 0:
state = self.env.reset()
state, self.stacked_frames = stack_frames(self.stacked_frames, state, True)
# Get next state, the rewards, done by taking a random action
choice = random.randint(1, len(self.possible_actions)) - 1
action = self.possible_actions[choice]
next_state, reward, done, _ = self.env.step(choice)
# stack the frames
next_state, self.stacked_frames = stack_frames(self.stacked_frames, next_state, False)
# if the episode is finished (we're dead)
if done:
# we inished the episode
next_state = np.zeros(state.shape)
# add experience to memory
self.memory.add((state, action, reward, next_state, done))
# start a new episode
state = self.env.reset()
state, self.stacked_frames = stack_frames(self.stacked_frames, state, True)
else:
# add experience to memory
self.memory.add((state, action, reward, next_state, done))
# our new state is now the next_state
state = next_state
# saver will help us save our model
self.saver = tf.train.Saver()
# setup tensorboard writer
self.writer = tf.summary.FileWriter("logs/")
# losses
tf.summary.scalar("Loss", self.DQNetwork.loss)
self.write_op = tf.summary.merge_all()
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = DQNetwork(state_size, action_size,
seed).to(device)
self.qnetwork_target = DQNetwork(state_size, action_size,
seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add(state, action, reward, next_state, done)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Variable]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# Get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(1)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
# Exploration parameters for epsilon greedy strategy
explore_start = 1.0 # exploration probability at start
explore_stop = 0.01 # minimum exploration probability
decay_rate = 0.0001 # exponential decay rate for exploration prob
# Q learning hyperparameters
discount_rate = 0.95 # Discounting rate
### MEMORY HYPERPARAMETERS
pretrain_length = batch_size # Number of experiences stored in the Memory when initialized for the first time
memory_size = 100 # Number of experiences the Memory can keep
tf.reset_default_graph()
DQNetwork = DQNetwork(state_size, action_size, learning_rate)
# PART II: GEN MEMORY
print("gen memory")
# class Memory():
# def __init__(self, max_size):
# self.buffer = deque(maxlen = max_size)
# def add(self, experience):
# self.buffer.append(experience)
# def sample(self, batch_size):
# buffer_size = len(self.buffer)
# index = np.random.choice(np.arange(buffer_size),
# size = batch_size,
class YellowBananaThief:
""" A smart agent that interacts with the environment to pick up yellow bananas"""
def __init__(self,
state_size,
action_size,
seed=0,
buffer_size=100000,
batch_size=64,
update_frequency=2,
gamma=.99,
learning_rate=5e-4,
tau=1e-3):
self.state_size = state_size
self.action_size = action_size
self.random = random.seed(seed)
self.batch_size = batch_size
self.memory = ReplayBuffer(self.action_size, buffer_size, batch_size,
seed)
self.time_step = 0
self.update_frequency = update_frequency
self.qnetwork_local = DQNetwork(state_size, action_size,
seed).to(device)
self.qnetwork_target = DQNetwork(state_size, action_size,
seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(),
lr=learning_rate)
# hyper-parameters
self.gamma = gamma
self.tau = tau
def act(self, state, epsilon):
""" Returns an epsilon greedy action to take in the current state
:param state: The current state in the environment
:param epsilon: Epsilon value to apply epsilon-greedy action selection
"""
def action_probabilities(action_vals, eps, num_actions):
""" Determine the epsilon probabilities of choosing actions """
probs = np.ones(num_actions, dtype=float) * (eps / num_actions)
best_action = np.argmax(action_vals)
probs[best_action] += (1. - eps)
return probs
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval(
) # get the network in evaluation mode and pull values from it
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train() # get the network back into train mode
action_probs = action_probabilities(action_values.cpu().data.numpy(),
epsilon, self.action_size)
return np.random.choice(np.arange(self.action_size), p=action_probs)
def step(self, state, action, reward, next_state, done):
""" Step forward to train the model """
self.memory.add(state, action, reward, next_state, done)
self.time_step = (self.time_step + 1) % self.update_frequency
if self.time_step == 0:
if len(self.memory) > self.batch_size:
# enough samples have been collected for learning from experience
experiences = self.memory.sample()
self.learn(experiences)
def learn(self, experiences):
""" Train the agent from a sample of experiences """
def soft_update(local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model (PyTorch model): weights will be copied from
target_model (PyTorch model): weights will be copied to
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
states, actions, rewards, next_states, dones = experiences
# max predicted Q values for the next state
q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(1)
# Q targets for current state
q_targets = rewards + (self.gamma * q_targets_next * (1 - dones))
# get expected q values from local model
q_expected = self.qnetwork_local(states).gather(1, actions)
# compute model loss
loss = F.mse_loss(q_expected, q_targets)
# minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
soft_update(self.qnetwork_local, self.qnetwork_target, self.tau)
def local_qnet(self):
""" Returns the trained model """
return self.qnetwork_local
import tensorflow as tf
import numpy as np
from environment import Environment
from model import DQNetwork
env = Environment()
DQNetwork = DQNetwork(learning_rate=0)
with tf.Session() as sess:
total_test_rewards = []
saver = tf.train.Saver()
saver.restore(sess, "./models/model.ckpt")
for episode in range(1):
total_rewards = 0
state = env.reset()
done = False
while not done:
state = state.reshape(