This repository contains JAX code for the implementation of the Deep Covering Eigenoptions (DCEO) algorithm which learns a diverse set of task-agnostic options to improve exploration.
We additionally provide a general guide to the integration of Laplacian-based options in a different codebase.
Deep Laplacian-based Options for Temporally-Extended Exploration
by Martin Klissarov and Marlos C. Machado.
DCEO is based on the idea of the Representation Driven Option Disovery cycle where options and representations are iteratively refined and bootstrapped from eachother. In this work, we argue that the Laplacian representation (also referred to as Proto-Value Functions and closely related to the Successor Representation) is very well-suited scaffold for option discovery as it naturally encodes the topology of the environment at various timescales.
We present a more in-depth overview in the project website and the blogpost.
In this repository, we share the code with respect to the Montezuma's Revenge experiments, which are built on the Dopamine codebase. To replicate the results on Montezuma's Revenge, only two files need to be added with respect to the original repository: full_rainbow_dceo.gin and full_rainbow_dceo.py which are both located in dopamine/jax/agents/full_rainbow.
For the sake of simplicity we include the complete Dopamine code source. To replicate results on Montezuma's Revenge, simply run the following
python -um dopamine.discrete_domains.train --base_dir results_folder \
--gin_files "dopamine/jax/agents/full_rainbow/configs/full_rainbow_dceo.gin" \
--gin_bindings "atari_lib.create_atari_environment.game_name='MontezumaRevenge'" \
--gin_bindings "JaxFullRainbowAgentDCEO.seed=1337"
The experiments should take from 5 to 7 days in order to run the complete 200M timesteps of training using one A100 GPU, 10 CPU cores and 60G of RAM.
Installation
-
To run experiments on Atari, you will need to get the Atari ROMS as described in the ale-py (i.e. by using
ale-import-romon directory containing the ROMS to import them), or by using atari-py (i.e.python -m atari_py.import_roms folder_containing_roms). -
To install the necessary requirements, with a
virutalenvor conda environment, simply do
pip install -r requirements.txt
- Finally, install Dopamine using the editable mode
pip install -e .
As DCEO performs remarkably well compared to several HRL baseline (such as DIAYN, CIC and DCO) as well as exploration baselines (such as RND and Counts), we believe it is important to facilitate the usage of Laplacian-based options. Therefore we must answer, what is the minimum amount of code needed to launch some experiments using such options? We answer this question by pointing at snippets of code in this repository.
Learning the Laplacian representation
The first step is to integrate the loss for learning the Laplacian representation on which the options are based. This can be done with the following code.
However, note that recent improvements on this representation learning method have lead to the Proper Laplacian for which a GitHub repository exists. The Proper Laplacian is much more robust and is better motivated theoretically.
square_div_dim = lambda x : x**2 / rep_dim
norm_transform = jnp.log1p if self._log_transform else square_div_dim # Defaults to jnp.log1p
def neg_loss_fn(phi_u, phi_v):
loss = 0
for dim in range(rep_dim, 0, -1):
coeff = coeff_vector[dim-1] - coeff_vector[dim]
u_norm = jnp.sqrt(jnp.dot(phi_u[:dim], phi_u[:dim]))
v_norm = jnp.sqrt(jnp.dot(phi_v[:dim], phi_v[:dim]))
dot_product = jnp.dot(phi_u[:dim], phi_v[:dim])
loss += coeff * (
dot_product ** 2 - norm_transform(u_norm) - norm_transform(v_norm) )
return loss
neg_loss_vmap = jax.vmap(neg_loss_fn)
def train_rep(rep_params, optimizer, optimizer_state, states):
def loss_fn(params):
"""Calculates loss given network parameters and transitions."""
def rep_online(state):
return rep_network_def.apply(params, state) # Same CNN+MLP encoder as for the Q values
phis = jax.vmap(rep_online)(states).output
phis = jnp.split(phis, 4, axis=0)
phi_current, phi_next, phi_random_u, phi_random_v = phis[0], phis[1], phis[2], phis[3]
pos_loss = ((phi_current - phi_next)**2).dot(coeff_vector[:rep_dim])
neg_loss = neg_loss_vmap(phi_random_u, phi_random_v)
loss = pos_loss + neg_loss
loss = jnp.mean(loss)
return loss, (jnp.mean(pos_loss), jnp.mean(neg_loss))
grad_fn = jax.grad(loss_fn, has_aux=True)
grad, (pos_loss, neg_loss) = grad_fn(rep_params)
updates, optimizer_state = optimizer.update(grad, optimizer_state,
params=rep_params)
rep_params = optax.apply_updates(rep_params, updates)
return optimizer_state, rep_params, pos_loss, neg_loss
self._train_rep = jax.jit(train_rep, static_argnames=('optimizer'))
The optimization can then be done by sampling state from a buffer and calling _train_rep
### Train the representation module ###
phi_tm1, phi_t, phi_u, phi_v = self._rep_sample_from_replay_buffer()
all_phis = (phi_tm1, phi_t, phi_u, phi_v)
(self.rep_optimizer_state, self.rep_params,
pos_loss, neg_loss) = self._train_rep(
self.rep_params, self.optimizer,
self.rep_optimizer_state, all_phis)
Here phi_current and phi_next are consecutive states whereas phi_random_u and phi_random_u are randomly sampled states.
Option Learning
To learn each of the options, we need to define the intrinsic reward from Section 3 in the paper. We then iterate over all the options and update each of them using this intrinsic reward signal.
for o in onp.random.choice(self.num_options, 1, replace=False):
option = self.options[o]
self._sample_from_replay_buffer()
current_states = self.preprocess_fn(self.replay_elements['state'])
next_states = self.preprocess_fn(self.replay_elements['next_state'])
all_states = onp.vstack((current_states, next_states))
rep = jax.vmap(self._get_rep, in_axes=(None, 0))(
self.rep_params, all_states).q_values
rep = onp.asarray(rep)
# Calculate the Laplacian representation between two consecutive states
# and index the representation by the current option being updated.
current_laplacian_o = rep[:len(rep) // 2, o]
next_laplacian_o = rep[len(rep) // 2:, o]
intr_rew = next_laplacian_o - current_laplacian_o # Line 155 in the paper
# Usual Q-Learning, or your preferred RL method
Option Execution
Our option execution algorithm defined in Algorithm 1 in the paper is straightforward as it does not rely on meta-policies or parametrized terminations functions. The following code implements it.
option_term = option_term or onp.random.rand() < (1 / dur)
epsilon = jnp.where(
eval_mode, epsilon_eval,
epsilon_fn(epsilon_decay_period, training_steps, min_replay_history,
epsilon_train))
if option_term:
cur_opt = None
if onp.random.rand() < epsilon:
if onp.random.rand() < option_prob:
cur_opt = onp.random.randint(num_options)
option_term = False
params = main_policy_params if cur_opt is None else options[cur_opt].online_params
rng, action = act(
network_def, params, state, rng, num_actions, eval_mode, support, epsilon)
Hopefully this should make it easier to implement Laplacian-based options.
If you found this repository useful, we invite you to cite our work.
@inproceedings{klissarov2023deep,
title = {Deep Laplacian-based Options for Temporally-Extended Exploration},
author = {Martin Klissarov and Marlos C. Machado},
booktitle = {International Conference on Machine Learning (ICML)},
year = {2023}
}