/lucid-dreaming-for-exp-replay

Implementation of paper "Lucid Dreaming for Experience Replay: Refreshing Past States with the Current Policy"

Primary LanguagePythonMIT LicenseMIT

Lucid Dreaming for Experience Replay (LiDER)

Implementation of our paper Lucid Dreaming for Experience Replay: Refreshing Past States with the Current Policy

Tested in Ubuntu 18.04, Tensorflow 1.11, Python 3.6

Citation:

@article{Du2020LiDER,
  title={Lucid Dreaming for Experience Replay: Refreshing Past States with the Current Policy},
  author={Yunshu Du and Garrett Warnell and A. Gebremedhin and P. Stone and Matthew E. Taylor},
  journal={ArXiv},
  year={2020},
  volume={abs/2009.13736}
}

Available in this repo

A3CTBSIL, LiDER, three LiDER ablations, LiDER-TA, LiDER-BC

Data and plots, analyses

Installation

It is recommended to use a virtual environment for installation. Using conda (version 4.7.10) as an example here.

Create a virtual environment (name "liderenv") from the env.yml file:

$ conda env create -f env.yml

Activate the environment

$ conda activate liderenv

This "liderenv" environment installs a CPU-only TensorFlow version tensorflow==1.11 to avoid the complexity of setting up CUDA (our implementation can run in either CPU or GPU).

If you have a supported GPU and have already set up CUDA, install tensorflow-gpu by replacing tensorflow==1.11.0 with tensorflow-gpu==1.11.0 in the env.yml file. For more informatioon about GPU suppoert, see Tensorflow website.

Note: you may use a newer version of Tensorflow (e.g., 1.14) but there will be warning messages. However, this implementation has not been tested in TensorFlow 1.14 or higher

How to run

We provide several bash files that have been configured to run corresponding experiments in the paper (i.e., using the same parameters as in the paper). The bash file takes the first argument as the input to choose which game to train. Valid games are: Gopher, NameThisGame, Alien, MsPacman, Freeway, and MontezumaRevenge.

Note: If you installed tensorflow-gpu, uncomment line --use-gpu --cuda-devices=0 in the bash file to enable running on GPU.

A3CTBSIL, LiDER, and LiDER ablation studies

For example, to train MsPacman in baseline A3CTBSIL:

$./run_a3ctbsil.sh MsPacman

To run LiDER, our proposed framework:

$./run_lider.sh MsPacman

To run the ablation LiDER-AddAll:

$./run_lider_addall.sh MsPacman

To run the ablation LiDER-OneBuffer:

$./run_lider_onebuffer.sh MsPacman

To run the ablation LiDER-SampleR:

$./run_lider_sampleR.sh MsPacman

LiDER-TA and LiDER-BC

First, download pretrained_models.tar.gz (file size ~2GB). Then, unzip to obtain all pre-trained models used in the paper:

$ tar -xzvf pretrained_models.tar.gz

This will produce folder pretrained_models, which contains the checkpoints for all pre-trained models. This folder should be placed under the main directory:

lucid-dreaming-for-exp-replay
└──LiDER
└──pretrained_models
   └── TA
        └── Alien
        └── Freeway
        └── Gopher
        └── MontezumaRevenge
        └── MsPacman
        └── NameThisGame
   └── BC
        └── Alien
        └── Freeway
        └── Gopher
        └── MontezumaRevenge
        └── MsPacman
        └── NameThisGame

Then, to run the extension LiDER-TA:

$./run_lider_ta MsPacman

To run the extension LiDER-BC:

$./run_lider_bc MsPacman

Checkpointing

By default, our implementation checkpoints the training process every 1 million training steps and dumps results to the results folder (which will be created if not already exists), including:

  • Current model parameters: folder model_checkpoints

  • Best model parameters: folder model_best

  • Testing rewards: pickle file with pattern [gamename]_v4-a3c-rewards.pkl)

  • Additional statistics used for analyses, see Section 4.2 of the paper: pickle files with pattern [gamename]_v4-a3c-dict-*.pkl

  • Replay buffers: pickle files with pattern [gamename]_v4-*mem-*.pkl

At the beginning of a run, our implementation looks for whether a folder for the same experiment setting already exists for the current experiment. If so, it will resume from the last saved checkpoint; otherwise, it will create a new experiment folder and start training from scratch.

The experiment folder will be created under results during training based on the current setting. For example. the experiment folder for training LiDER on MsPacman will be named MsPacmanNoFrameskip_v4_rawreward_transformedbell_sil_prioritymem_lider_[trial_number].

[trial_number] is set by parameter --append-experiment-num in the bach file. By default, --append-experiment-num=1. To run more trials, set it to a different number, such as --append-experiment-num=2 or --append-experiment-num=3.

The checkpointing frequency can be changed by modifying argument --checkpoint-freq=1 to any number between 0 and 50. For example, --checkpoint-freq=5 means to checkpoint every 5 million training steps.

The argument --checkpoint-buffer controls whether to save the entire replay buffer. When enabled, the replay buffer(s) will be saved at each checkpoint; previously collected experiences will be loaded when training is resumed.

Note that saving replay buffers can be memory-intensive. If you don't have enough capacity, you can either increase --checkpoint-freq to checkpoint less frequently, or remove line --checkpoint-buffer to save only the model parameters. If replay buffers were not saved, when resuming training you will start with empty buffer(s).

Data and plots

Reproduce figures and analyses in the paper

See folder data_and_plots.

Plot your experiment results

When running your own experiments, a results folder will be created to save results. For example, after training MsPacman using LiDER for three trials, the results folder will structure as follows:

lucid-dreaming-for-exp-replay
└──LiDER
└──pretrained_models
└──results
   └── a3c
        └── MsPacmanNoFrameskip_v4_rawreward_transformedbell_sil_prioritymem_lider_1
        └── MsPacmanNoFrameskip_v4_rawreward_transformedbell_sil_prioritymem_lider_2
        └── MsPacmanNoFrameskip_v4_rawreward_transformedbell_sil_prioritymem_lider_3
   └── log
        └── a3c
            └── MsPacmanNoFrameskip_v4

results/a3c saves model checkpoints and testing rewards (see descriptions under Checkpointing). results/log saves tensorboard events so that you can monitor the training process. To launch tensorboard, follow the standard procedure and set --logdir to corresponding log folders. See TensorBoard website for more instructions.

The .pkl files saved under results/a3c/[experiment_name] are used for plotting the testing results. Use the plotting script generate_plots.py to generate plots. For example, to plot A3CTBSIL and LiDER:

$ python3 generate_plots.py --env=MsPacman --baseline --lidera3c --saveplot

Here is a list of available parameters in the plotting script:

  • Parameters for general settings:

    • --env: which game to plot. For example, --env=MsPacman.
    • --max-timesteps: the number of time steps (in million) to plot. For example,--max-timesteps=30 will plot results for 30 million steps. The default value is 50 million steps.
    • --saveplot: when enabled, save the plot; otherwise, only display the plot without saving. The default value is False (i.e., not saving).
    • --nolegend: when enabled, no legend is shown. The default value is False (i.e., show legend).
    • --folder: where to save the plot. By default, a new folder plots will be created.
    • --fn: file name of the plot. When --saveplot is enabled, --fn must be provided to save the plot.

  • Parameters for plotting each algorithm:

    • --baseline: A3CTBSIL
    • --lidera3c: LiDER
    • --addall: LiDER-AddAll
    • --onebuffer: LiDER-OneBuffer
    • --sampler: LiDER-SampleR
    • --liderta: LiDER-TA
    • --liderbc: LiDER-BC
Note on plotting LiDER-TA and LiDER-BC

The values of the horizontal lines showing pretrained TAs and BCs' performance (Figure 7 of our paper) need to be supplied manually. The current values are taken from our paper. When running your own experiments, the TA and BC's performance will be evaluated for 50 episodes at the beginning of training. Their evaluation results will be stored under pretrained_models/TA (or BC)/[game]/[game]-model-eval.txt, including the episodic mean reward, the standard deviation, and the reward for each episode.

Plot your analyses

Analyses can be generated for your A3CTBSIL and LiDER experiments (currently not supporting analysis for other extensions). The results are also stored under the results folder. Use script generate_analyses.py to generate analyses plot. For example, to look at the "success rate" in refresher worker (Section 4.2.1):

$ python3 generate_analyses.py --env=MsPacman --refresh-success --saveplot

Here is a list of available parameters in the plotting script:

  • Parameters for general settings are the same as above

  • Parameters for plotting each analysis:

    • Section 4.2.1:
      • --refresh-success: Success rate
      • --refresh-gnew: Gnew vs. G
    • Section 4.2.2:
      • --sil-old-used: Old samples used (this does not generate plots, only displays values of the table)
      • --batch-sample-usage-ratio: Batch sample usage ratio (LiDER)
      • --sil-sample-usage-ratio: SIL sample usage ratio
      • --sil-return-of-used-samples: Return of used samples (LiDER)
    • Section 4.2.3:
      • --total-batch-usage: batch sample usage ratio (A3CTBSIL & LiDER)
      • --total-used-return: Return of used samples (A3CTBSIL & LiDER)

Acknowledgements

We thank Gabriel V. de la Cruz Jr. for helpful discussions; his open-source code at github.com/gabrieledcjr/DeepRL is used for training the behavior cloning models in this work. This research used resources of Kamiak, Washington State University’s high-performance computing cluster. Assefaw Gebremedhin is supported by the NSF award IIS-1553528. Part of this work has taken place in the Intelligent Robot Learning (IRL) Lab at the University of Alberta, which is supported in part by research grants from the Alberta Machine Intelligence Institute (Amii), CIFAR, and NSERC. Part of this work has taken place in the Learning Agents Research Group (LARG) at UT Austin. LARG research is supported in part by NSF (CPS-1739964, IIS1724157, NRI-1925082), ONR (N00014-18-2243), FLI (RFP2-000), ARL, DARPA, Lockheed Martin, GM, and Bosch. Peter Stone serves as the Executive Director of Sony AI America and receives financial compensation for this work. The terms of this arrangement have been reviewed and approved by the University of Texas at Austin in accordance with its policy on objectivity in research.