search

takuseno / d3rlpy-benchmarks

Benchmark data for d3rlpy
MIT License
20 stars 5 forks source link

lack of Ant-v2 task related #7

Open im-Kitsch opened 10 months ago

im-Kitsch commented 10 months ago

Hello,

I am wondering learning results of Ant-v2's dataset, like ant-random-v0/2 ant-medium-v0/v2. I think it is not listed in d4RL original paper but it's supplemented in github later.

Do you have the interest to evaluate thoes ant-* environment? Cuz I get very strange result for ant, even ant-random gives great result. So I am curious to compare with your result.

Best,

takuseno commented 10 months ago

@im-Kitsch Hi, thanks for the issue. It takes long to benchmark all algorithms with Ant dataset. But, if you give me which algorithm you want me to test, I can prioritize it.

im-Kitsch commented 10 months ago

Hi, thnks a lot for the kind replay. Personally I am interesting on AWAC or BCQ.

Thanks!

takuseno commented 10 months ago

Okay, I started BCQ benchmark with ant-x-v0 dataset first. I'll be back to you later once I get results.

takuseno commented 9 months ago

@im-Kitsch Sorry, it took time, but I've added BCQ results with ant-x-v0 datasets. https://github.com/takuseno/d3rlpy-benchmarks/blob/main/d4rl_table.csv

takuseno commented 9 months ago

I've added v2 dataset results as well.

im-Kitsch commented 9 months ago

cool, @takuseno , really thanks a lot, it's really helpful