search

ermongroup / multiagent-gail

MIT License
80 stars 24 forks source link

Parameters for expert policy #4

Open wsjeon opened 5 years ago

wsjeon commented 5 years ago

Dear authors,

Hi. Thank you for sharing your codes.

Recently, I've been interested in MAGAIL and tried to reproduce your results. Firstly, I tried to train expert policy as recommended in README.md by python -m sandbox.mack.run_simple, but I failed. I thought there is a problem on action dimension, so I made all actions as multi-hot vectors and modified relevant terms. After training with MACK, however, it seems like agents cannot recover the appropriate policies similar to MADDPG.

So I wonder whether it is possible to share the weight files of expert so that readers can simply generate expert trajectories.

Thanks.

Ericonaldo commented 5 years ago

Also have a problem on action dimension when running codes follows README.md, with some private path problems...

kvas7andy commented 4 years ago

@wsjeon Hi! May I ask if you eventually came up with weights with good performance (for expert trajectories with MACK)?

I am currently running the cooperative navigarion _"simple__spread" environment, and all losses rise at some point

kvas7andy commented 4 years ago

@wsjeon @Ericonaldo An update for the code: I found new implementation of MAGAIL used for the new paper of ermongroup here. The paper is "Multi-Agent Adversarial Inverse Reinforcement Learning". Results of execution are similar to new paper, but I cannot reach the results of MAGAIL paper (based on the paper "Multi-Agent Generative Adversarial Imitation Learning").

Anyone reached performance from this paper with MAGAIL?