Atari 2600: Pong with DQN (type-I q-function)

In this notebook we solve the Pong environment using the classic DQN agent. We’ll use a convolutional neural net (without pooling) as our function approximator for the q-function.

This version differs from the standard DQN version in that the q-function is modelled as a type-I q-function, i.e.

\[(s, a)\ \mapsto\ q(s,a) \in \mathbb{R}\]

instead of the standard (type-II):

\[s\ \mapsto\ q(s, .) \in \mathbb{R}^n\]

This notebook periodically generates GIFs, so that we can inspect how the training is progressing.

After a few hundred episodes, this is what you can expect:

Beating Atari 2600 Pong after a few hundred episodes.

Open in Google Colab
import os

# set some env vars
os.environ.setdefault('JAX_PLATFORM_NAME', 'gpu')     # tell JAX to use GPU
os.environ['XLA_PYTHON_CLIENT_MEM_FRACTION'] = '0.1'  # don't use all gpu mem
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'              # tell XLA to be quiet

import gymnasium
import jax
import coax
import haiku as hk
import jax.numpy as jnp
from optax import adam

# the name of this script
name = 'dqn_type1'

# env with preprocessing
env = gymnasium.make('PongNoFrameskip-v4', render_mode='rgb_array')  # AtariPreprocessing will do frame skipping
env = gymnasium.wrappers.AtariPreprocessing(env)
env = coax.wrappers.FrameStacking(env, num_frames=3)
env = gymnasium.wrappers.TimeLimit(env, max_episode_steps=108000 // 3)
env = coax.wrappers.TrainMonitor(env, name=name, tensorboard_dir=f"./data/tensorboard/{name}")

def func(S, A, is_training):
    """ type-1 q-function: (s,a) -> q(s,a) """
    body = hk.Sequential((
        hk.Conv2D(16, kernel_shape=8, stride=4), jax.nn.relu,
        hk.Conv2D(32, kernel_shape=4, stride=2), jax.nn.relu,
    head = hk.Sequential((
        hk.Linear(256), jax.nn.relu,
        hk.Linear(1, w_init=jnp.zeros), jnp.ravel
    X = jnp.stack(S, axis=-1) / 255.  # stack frames
    return head(jax.vmap(jnp.kron)(body(X), A))

# function approximator
q = coax.Q(func, env)
pi = coax.BoltzmannPolicy(q, temperature=0.015)  # <--- different from standard DQN

# target network
q_targ = q.copy()

# updater
qlearning = coax.td_learning.QLearning(q, q_targ=q_targ, optimizer=adam(3e-4))

# reward tracer and replay buffer
tracer = coax.reward_tracing.NStep(n=1, gamma=0.99)
buffer = coax.experience_replay.SimpleReplayBuffer(capacity=1000000)

while env.T < 3000000:
    s, info = env.reset()

    for t in range(env.spec.max_episode_steps):
        a = pi(s)
        s_next, r, done, truncated, info = env.step(a)

        # trace rewards and add transition to replay buffer
        tracer.add(s, a, r, done or truncated)
        while tracer:

        # learn
        if len(buffer) > 50000:  # buffer warm-up
            metrics = qlearning.update(buffer.sample(batch_size=32))

        if env.T % 10000 == 0:
            q_targ.soft_update(q, tau=1)

        if done or truncated:

        s = s_next

    # generate an animated GIF to see what's going on
    if env.period(name='generate_gif', T_period=10000) and env.T > 50000:
        T = env.T - env.T % 10000  # round to 10000s
            env=env, policy=pi, resize_to=(320, 420),