Any plan for providing reinforcement learning?

[lifelongeek]

As far as I know, ParlAI does not provide reinforcement learning dialog example for now.

Anyone considering to commit it? If so, could you share some info such as

• reference paper
• expected schedule

[jaseweston]

There are no RL code examples right now, but ParlAI is set up to use RL. E.g. we included a reward field in the action/observation message... I think these datasets in ParlAI use it (but it's fixed, so more imitation learning): DBLL-bAbI and DBLL-Movie

[lifelongeek]

Thanks. Could you provide some detail info of DBLL-bAbI & DBLL-Movie?

reference document (probably the paper "Dialog based Language learning"?) example script (probably : python examples/memnn_luatorch_cpu/full_task_train.py -t dbll_movie -bs 32 )

[jaseweston]

On Thu, Aug 3, 2017 at 12:56 AM, Geonmin Kim notifications@github.com wrote:

Thanks. Could you provide some detail info of DBLL-bAbI & DBLL-Movie?

reference document (probably the paper "Dialog based Language learning"?)

Yes, in general you can see dataset details here: https://github.com/facebookresearch/ParlAI/blob/master/parlai/tasks/task_list.py

example script (probably : python examples/memnn_luatorch_cpu/full_task_train.py

-t dbll_movie -bs 32 )

you could use any model potentially yes

python examples/train_model.py -m seq2seq -t dbll-babi -bs 8 -mf /tmp/model_s2s

but we don't have any models yet that use rewards or forward prediction (please add if you like! and we are always adding more, so more soon..)

—

You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/facebookresearch/ParlAI/issues/244#issuecomment-319867721, or mute the thread https://github.com/notifications/unsubscribe-auth/AKjk-Ef7O7vyz0yOFgzyBneJiksXl0Q0ks5sUVLrgaJpZM4On9wJ .