Skip to content
master
Switch branches/tags
Go to file
Code

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

RLlib Reference Results

Benchmarks of RLlib algorithms against published results. These benchmarks are a work in progress. For other results to compare against, see yarlp and more plots from OpenAI.

Ape-X Distributed Prioritized Experience Replay

rllib train -f atari-apex/atari-apex.yaml

Comparison of RLlib Ape-X to Async DQN after 10M time-steps (40M frames). Results compared to learning curves from Mnih et al, 2016 extracted at 10M time-steps from Figure 3.

env RLlib Ape-X 8-workers Mnih et al Async DQN 16-workers Mnih et al DQN 1-worker
BeamRider 6134 ~6000 ~3000
Breakout 123 ~50 ~10
QBert 15302 ~1200 ~500
SpaceInvaders 686 ~600 ~500

Here we use only eight workers per environment in order to run all experiments concurrently on a single g3.16xl machine. Further speedups may be obtained by using more workers. Comparing wall-time performance after 1 hour of training:

env RLlib Ape-X 8-workers Mnih et al Async DQN 16-workers Mnih et al DQN 1-worker
BeamRider 4873 ~1000 ~300
Breakout 77 ~10 ~1
QBert 4083 ~500 ~150
SpaceInvaders 646 ~300 ~160

Ape-X plots: apex

IMPALA and A2C

rllib train -f atari-impala/atari-impala.yaml

rllib train -f atari-a2c/atari-a2c.yaml

RLlib IMPALA and A2C on 10M time-steps (40M frames). Results compared to learning curves from Mnih et al, 2016 extracted at 10M time-steps from Figure 3.

env RLlib IMPALA 32-workers RLlib A2C 5-workers Mnih et al A3C 16-workers
BeamRider 2071 1401 ~3000
Breakout 385 374 ~150
QBert 4068 3620 ~1000
SpaceInvaders 719 692 ~600

IMPALA and A2C vs A3C after 1 hour of training:

env RLlib IMPALA 32-workers RLlib A2C 5-workers Mnih et al A3C 16-workers
BeamRider 3181 874 ~1000
Breakout 538 268 ~10
QBert 10850 1212 ~500
SpaceInvaders 843 518 ~300

IMPALA plots: tensorboard

A2C plots: tensorboard

Pong in 3 minutes

With a bit of tuning, RLlib IMPALA can solve Pong in ~3 minutes:

rllib train -f pong-speedrun/pong-impala-fast.yaml

tensorboard

DQN / Rainbow

rllib train -f atari-dqn/basic-dqn.yaml rllib train -f atari-dqn/duel-ddqn.yaml rllib train -f atari-dqn/dist-dqn.yaml

RLlib DQN after 10M time-steps (40M frames). Note that RLlib evaluation scores include the 1% random actions of epsilon-greedy exploration. You can expect slightly higher rewards when rolling out the policies without any exploration at all.

env RLlib Basic DQN RLlib Dueling DDQN RLlib Distributional DQN Hessel et al. DQN Hessel et al. Rainbow
BeamRider 2869 1910 4447 ~2000 ~13000
Breakout 287 312 410 ~150 ~300
QBert 3921 7968 15780 ~4000 ~20000
SpaceInvaders 650 1001 1025 ~500 ~2000

Basic DQN plots: tensorboard

Dueling DDQN plots: tensorboard

Distributional DQN plots: tensorboard

Proximal Policy Optimization

rllib train -f atari-ppo/atari-ppo.yaml

rllib train -f halfcheetah-ppo/halfcheetah-ppo.yaml

2018-09:

RLlib PPO with 10 workers (5 envs per worker) after 10M and 25M time-steps (40M/100M frames). Note that RLlib does not use clip parameter annealing.

env RLlib PPO @10M RLlib PPO @25M Baselines PPO @10M
BeamRider 2807 4480 ~1800
Breakout 104 201 ~250
QBert 11085 14247 ~14000
SpaceInvaders 671 944 ~800

tensorboard

RLlib PPO wall-time performance vs other implementations using a single Titan XP and the same number of CPUs. Results compared to learning curves from Fan et al, 2018 extracted at 1 hour of training from Figure 7. Here we get optimal results with a vectorization of 32 environment instances per worker:

env RLlib PPO 16-workers Fan et al PPO 16-workers TF BatchPPO 16-workers
HalfCheetah 9664 ~7700 ~3200

tensorboard

2020-01:

Same as 2018-09, comparing only RLlib PPO-tf vs PPO-torch.

env RLlib PPO @20M (tf) RLlib PPO @20M (torch) plot
BeamRider 4142 3850 tensorboard
Breakout 132 166 tensorboard
QBert 7987 14294 tensorboard
SpaceInvaders 956 1016 tensorboard

Soft Actor Critic

rllib train -f halfcheetah-sac/halfcheetah-sac.yaml

RLlib SAC after 3M time-steps.

RLlib SAC versus SoftLearning implementation Haarnoja et al, 2018 benchmarked at 500k and 3M timesteps respectively.

env RLlib SAC @500K Haarnoja et al SAC @500K RLlib SAC @3M Haarnoja et al SAC @3M
HalfCheetah 9000 ~9000 13000 ~15000

tensorboard

MAML

MAML uses additional metrics to measure performance; episode_reward_mean measures the agent's returns before adaptation, episode_reward_mean_adapt_N measures the agent's returns after N gradient steps of inner adaptation, and adaptation_delta measures the difference in performance before and after adaptation.

rllib train -f maml/halfcheetah-rand-direc-maml.yaml

tensorboard

rllib train -f maml/ant-rand-goal-maml.yaml

tensorboard

rllib train -f maml/pendulum-mass-maml.yaml

tensorboard

MB-MPO

rllib train -f mbmpo/halfcheetah-mbmpo.yaml

rllib train -f mbmpo/hopper-mbmpo.yaml

MBMPO uses additional metrics to measure performance. For each MBMPO iteration, MBMPO samples fake data from the transition dynamics workers and steps through MAML for N iterations. MAMLIter$i$_DynaTrajInner_$j$_episode_reward_mean corresponds to agent's performance across the dynamics models at the ith iteration of MAML and the jth step of inner adaptation.

RLlib MBMPO versus Clavera et al, 2018 benchmarked at 100k timesteps. Results reported below were ran on RLLib and the master branch of the original codebase respectively.

env RLlib MBPO @100K Clavera et al MBMPO @100K
HalfCheetah 520 ~550
Hopper 620 ~650

tensorboard

Dreamer

rllib train -f dreamer/dreamer-deepmind-control.yaml

RLlib Dreamer at 1M time-steps.

RLlib Dreamer versus Google implementation Danijar et al, 2020 benchmarked at 100k and 1M timesteps respectively.

env RLlib Dreamer @100K Danijar et al Dreamer @100K RLlib Dreamer @1M Danijar et al Dreamer @1M
Walker 320 ~250 920 ~930
Cheetah 300 ~250 640 ~800

tensorboard

RLlib Dreamer also logs gifs of Dreamer's imagined trajectories (Top: Ground truth, Middle: Model prediction, Bottom: Delta).

Alt Text Alt Text

CQL

rllib train -f halfcheetah-cql/halfcheetah-cql.yaml

rllib train -f halfcheetah-cql/halfcheetah-bc.yaml

Since CQL is an offline RL algorithm, CQL's returns are evaluated only during the evaluation loop (once every 1000 gradient steps).

RLlib CQL versus Behavior Cloning (BC) benchmarked at 500K gradient steps over the dataset. Results reported below were ran on RLLib. The only difference between BC and CQL is the bc_iters parameter in CQL (how many iterations to run BC loss).

HalfCheetah-Random-v1 is a dataset collected by a random policy, with an average performance of around -400 reward. CQL can do better than the original policy's performance.

env RLlib BC @500K RLlib CQL @500K
HalfCheetah-Random-v1 -600 -85

tensorboard

Transformers

rllib train -f vizdoom-attention/vizdoom-attention.yaml

RLlib's model catalog feature implements a variety of different models for the policy and value network, one of which supports using attention in RL. In particular, RLlib implements a Gated Transformer (Parisotta et al, 2019), abbreviated as GTrXL.

GTrXL is benchmarked in the Vizdoom environment, where the goal is to shoot a monster as quickly as possible. With PPO as the algorithm and GTrXL as the model, RLlib can successfuly solve the Vizdoom environment and reach human level performance.

env RLlib Transformer @2M
VizdoomBasic-v0 ~75

tensorboard