fietelab / farcuriosity

Neuro-Inspired Fragmentation and Recall to Overcome Catastrophic Forgetting in Curiosity

Neuro-Inspired Fragmentation and Recall to Overcome Catastrophic Forgetting in Curiosity

Official implementation for Neuro-Inspired Fragmentation and Recall to Overcome Catastrophic Forgetting in Curiosity (NeurIPS 2023 Workshop).

Usage

Install packages.

pip install -r requirements.txt

Run Experiments.

python launch.py -alg ppo -curiosity_alg rnd -env jamesbond -lstm -sample_mode gpu -num_gpus 1 -normalize_advantage -normalize_reward -dual_value -normalize_obs -fragmentation -recall -use_feature -use_wandb

Citation

@inproceedings{hwang2023neuro,
    author = {Hwang, Jaedong and Hong, Zhang-Wei and Chen, Eric and Boopathy, Akhilan and Agrawal, Pulkit and Fiete, Ila},
    title = {Neuro-Inspired Fragmentation and Recall to Overcome Catastrophic Forgetting in Curiosity},
    booktitle={NeurIPSW 2023 (Intrinsically Motivated Open-ended Learning)},
    year = {2023},
}   

Acknowledgements

Our code is based on EIPO and rlpyt

farcuriosity's Issues

Running the default README command wouldn't go well

Hi.

I used the following command in order get the related results in the paper:

Run Experiments.

python launch.py -alg ppo -curiosity_alg rnd -env jamesbond -lstm -sample_mode gpu -num_gpus 1 -normalize_advantage -normalize_reward -dual_value -normalize_obs -fragmentation -recall -use_feature -use_wandb
However, the result is much more like the setting which doesn't use the FARCuriostiy algorithm.

I had an idea that what will happen if we remove the
-normalize_reward
argument, and it worked as it was expected.

Could you give me a hint on what is the problem with the default run command? Is it preferred to use it or this command is just a hint?