Soft Actor-Critic (SAC)

Action Space	Continuous
Paper	Soft Actor-Critic Algorithms and Applications [1]
Framework(s)	PyTorch
API Reference	garage.torch.algos.SAC
Code	garage/torch/algos/sac.py
Examples	examples

Soft Actor-Critic (SAC) is an algorithm which optimizes a stochastic policy in an off-policy way, forming a bridge between stochastic policy optimization and DDPG-style approaches. A central feature of SAC is entropy regularization. The policy is trained to maximize a trade-off between expected return and entropy, a measure of randomness in the policy. This has a close connection to the exploration-exploitation trade-off: increasing entropy results in more exploration, which can accelerate learning later on. It can also prevent the policy from prematurely converging to a bad local optimum.

Default Parameters

initial_log_entropy=0.
discount=0.99
buffer_batch_size=64
min_buffer_size=int(1e4)
target_update_tau=5e-3
policy_lr=3e-4
qf_lr=3e-4
reward_scale=1.0
optimizer=torch.optim.Adam
steps_per_epoch=1
num_evaluation_episodes=10

Examples

#!/usr/bin/env python3
"""This is an example to train a task with SAC algorithm written in PyTorch."""
import numpy as np
import torch
from torch import nn
from torch.nn import functional as F

from garage import wrap_experiment
from garage.envs import GymEnv, normalize
from garage.experiment import deterministic
from garage.replay_buffer import PathBuffer
from garage.sampler import LocalSampler
from garage.torch import set_gpu_mode
from garage.torch.algos import SAC
from garage.torch.policies import TanhGaussianMLPPolicy
from garage.torch.q_functions import ContinuousMLPQFunction
from garage.trainer import Trainer


@wrap_experiment(snapshot_mode='none')
def sac_half_cheetah_batch(ctxt=None, seed=1):
    """Set up environment and algorithm and run the task.

    Args:
        ctxt (garage.experiment.ExperimentContext): The experiment
            configuration used by Trainer to create the snapshotter.
        seed (int): Used to seed the random number generator to produce
            determinism.

    """
    deterministic.set_seed(seed)
    trainer = Trainer(snapshot_config=ctxt)
    env = normalize(GymEnv('HalfCheetah-v2'))

    policy = TanhGaussianMLPPolicy(
        env_spec=env.spec,
        hidden_sizes=[256, 256],
        hidden_nonlinearity=nn.ReLU,
        output_nonlinearity=None,
        min_std=np.exp(-20.),
        max_std=np.exp(2.),
    )

    qf1 = ContinuousMLPQFunction(env_spec=env.spec,
                                 hidden_sizes=[256, 256],
                                 hidden_nonlinearity=F.relu)

    qf2 = ContinuousMLPQFunction(env_spec=env.spec,
                                 hidden_sizes=[256, 256],
                                 hidden_nonlinearity=F.relu)

    replay_buffer = PathBuffer(capacity_in_transitions=int(1e6))

    sac = SAC(env_spec=env.spec,
              policy=policy,
              qf1=qf1,
              qf2=qf2,
              gradient_steps_per_itr=1000,
              max_episode_length_eval=1000,
              replay_buffer=replay_buffer,
              min_buffer_size=1e4,
              target_update_tau=5e-3,
              discount=0.99,
              buffer_batch_size=256,
              reward_scale=1.,
              steps_per_epoch=1)

    if torch.cuda.is_available():
        set_gpu_mode(True)
    else:
        set_gpu_mode(False)
    sac.to()
    trainer.setup(algo=sac, env=env, sampler_cls=LocalSampler)
    trainer.train(n_epochs=1000, batch_size=1000)


s = np.random.randint(0, 1000)
sac_half_cheetah_batch(seed=521)

References

1

Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, and others. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018.

---

This page was authored by Ruofu Wang (@yeukfu).