Conservative Q-Learning (CQL)#

CQL is an extension of Q-learning that addresses the typical overestimation of values induced by the distributional shift between the dataset and the learned policy in offline RL algorithms. A conservative Q-function is learned, such that the expected value of a policy under this Q-function lower-bounds its true value

CQL paper: https://arxiv.org/abs/2006.04779

Can I use it?#

	Action	Observation
Discrete	✔️	✔️
Continuous	✔️	✔️

So far, we have implemented CQN - CQL applied to DQN, which cannot be used on continuous action spaces. We will soon be adding other CQL extensions of algorithms for offline RL.

Example#

import gymnasium as gym
import h5py
from agilerl.components.replay_buffer import ReplayBuffer
from agilerl.algorithms.cqn import CQN

# Create environment and Experience Replay Buffer, and load dataset
env = gym.make('CartPole-v1')
try:
    state_dim = env.observation_space.n       # Discrete observation space
    one_hot = True                            # Requires one-hot encoding
except Exception:
    state_dim = env.observation_space.shape   # Continuous observation space
    one_hot = False                           # Does not require one-hot encoding
try:
    action_dim = env.action_space.n           # Discrete action space
except Exception:
    action_dim = env.action_space.shape[0]    # Continuous action space

channels_last = False # Swap image channels dimension from last to first [H, W, C] -> [C, H, W]

if channels_last:
    state_dim = (state_dim[2], state_dim[0], state_dim[1])

field_names = ["state", "action", "reward", "next_state", "done"]
memory = ReplayBuffer(action_dim=action_dim, memory_size=10000, field_names=field_names)
dataset = h5py.File('data/cartpole/cartpole_random_v1.1.0.h5', 'r')  # Load dataset

# Save transitions to replay buffer
dataset_length = dataset['rewards'].shape[0]
for i in range(dataset_length-1):
    state = dataset['observations'][i]
    next_state = dataset['observations'][i+1]
    if channels_last:
        state = np.moveaxis(state, [3], [1])
        next_state = np.moveaxis(next_state, [3], [1])
    action = dataset['actions'][i]
    reward = dataset['rewards'][i]
    done = bool(dataset['terminals'][i])
    memory.save2memory(state, action, reward, next_state, done)

agent = CQN(state_dim=state_dim, action_dim=action_dim, one_hot=one_hot)   # Create DQN agent

state = env.reset()[0]  # Reset environment at start of episode
while True:
    experiences = memory.sample(agent.batch_size)   # Sample replay buffer
    # Learn according to agent's RL algorithm
    agent.learn(experiences)

To configure the network architecture, pass a dict to the CQN net_config field. For an MLP, this can be as simple as:

NET_CONFIG = {
      'arch': 'mlp',      # Network architecture
      'h_size': [32, 32]  # Network hidden size
  }

Or for a CNN:

NET_CONFIG = {
      'arch': 'cnn',      # Network architecture
      'h_size': [128],    # Network hidden size
      'c_size': [32, 32], # CNN channel size
      'k_size': [8, 4],   # CNN kernel size
      's_size': [4, 2],   # CNN stride size
      'normalize': True   # Normalize image from range [0,255] to [0,1]
  }

agent = CQN(state_dim=state_dim, action_dim=action_dim, one_hot=one_hot, net_config=NET_CONFIG)   # Create CQN agent

Saving and loading agents#

To save an agent, use the saveCheckpoint method:

from agilerl.algorithms.cqn import CQN

agent = CQN(state_dim=state_dim, action_dim=action_dim, one_hot=one_hot)   # Create CQN agent

checkpoint_path = "path/to/checkpoint"
agent.saveCheckpoint(checkpoint_path)

To load a saved agent, use the load method:

from agilerl.algorithms.cqn import CQN

checkpoint_path = "path/to/checkpoint"
agent = CQN.load(checkpoint_path)

Parameters#

class agilerl.algorithms.cqn.CQN(state_dim, action_dim, one_hot, index=0, net_config={'arch': 'mlp', 'h_size': [64, 64]}, batch_size=64, lr=0.0001, learn_step=5, gamma=0.99, tau=0.001, mut=None, double=False, actor_network=None, device='cpu', accelerator=None, wrap=True)#

The CQN algorithm class. CQN paper: https://arxiv.org/abs/2006.04779

Parameters:

state_dim (list[int]) – State observation dimension
action_dim (int) – Action dimension
one_hot (bool) – One-hot encoding, used with discrete observation spaces
index (int, optional) – Index to keep track of object instance during tournament selection and mutation, defaults to 0
net_config (dict, optional) – Network configuration, defaults to mlp with hidden size [64,64]
batch_size (int, optional) – Size of batched sample from replay buffer for learning, defaults to 64
lr (float, optional) – Learning rate for optimizer, defaults to 1e-4
learn_step (int, optional) – Learning frequency, defaults to 5
gamma (float, optional) – Discount factor, defaults to 0.99
tau (float, optional) – For soft update of target network parameters, defaults to 1e-3
mut (str, optional) – Most recent mutation to agent, defaults to None
double (bool, optional) – Use double Q-learning, defaults to False
actor_network (nn.Module, optional) – Custom actor network, defaults to None
device (str, optional) – Device for accelerated computing, ‘cpu’ or ‘cuda’, defaults to ‘cpu’
accelerator (accelerate.Accelerator(), optional) – Accelerator for distributed computing, defaults to None
wrap (bool, optional) – Wrap models for distributed training upon creation, defaults to True

clone(index=None, wrap=True)#

Returns cloned agent identical to self.

Parameters:: index (int, optional) – Index to keep track of agent for tournament selection and mutation, defaults to None

getAction(state, epsilon=0, action_mask=None)#

Returns the next action to take in the environment. Epsilon is the probability of taking a random action, used for exploration. For epsilon-greedy behaviour, set epsilon to 0.

Parameters:

state (numpy.ndarray[float]) – State observation, or multiple observations in a batch
epsilon (float, optional) – Probablilty of taking a random action for exploration, defaults to 0
action_mask (numpy.ndarray, optional) – Mask of legal actions 1=legal 0=illegal, defaults to None

learn(experiences)#

Updates agent network parameters to learn from experiences.

Parameters:: experiences – List of batched states, actions, rewards, next_states, dones in that order.

classmethod load(path, device='cpu', accelerator=None)#

Creates agent with properties and network weights loaded from path.

Parameters:

path (string) – Location to load checkpoint from
device (str, optional) – Device for accelerated computing, ‘cpu’ or ‘cuda’, defaults to ‘cpu’
accelerator (accelerate.Accelerator(), optional) – Accelerator for distributed computing, defaults to None

loadCheckpoint(path)#

Loads saved agent properties and network weights from checkpoint.

Parameters:: path (string) – Location to load checkpoint from

saveCheckpoint(path)#

Saves a checkpoint of agent properties and network weights to path.

Parameters:: path (string) – Location to save checkpoint at

softUpdate()#: Soft updates target network.

test(env, swap_channels=False, max_steps=500, loop=3)#

Returns mean test score of agent in environment with epsilon-greedy policy.

Parameters:

env (Gym-style environment) – The environment to be tested in
swap_channels (bool, optional) – Swap image channels dimension from last to first [H, W, C] -> [C, H, W], defaults to False
max_steps (int, optional) – Maximum number of testing steps, defaults to 500
loop (int, optional) – Number of testing loops/episodes to complete. The returned score is the mean. Defaults to 3