Is there any method to add DQN to convlab2? about convlab-2 HOT 7 CLOSED

In multi-domain dialogue, multiple actions are common. DQN cannot solve this directly. In ConvLab, most frequent dialogue act combinations are considered as actions, which could not cover every possible dialogue act combinations. However, we welcome your contributions to adding more RL methods. You can follow the structure of the existing RL policy in ConvLab-2. Look forward to your success!

from convlab-2.

Hey, since my focus point is in RL-Reward function part, so for me there are not so much RL policies I could use, there are only PPO, and PG, is there any other RL methods I could implement in this platform?

I have made some statics, you are right, there are more than 8000 actions in this Multi Domain dataset.

from convlab-2.

You could try DQN, A3C, A2C, or any other that you are interested in.

from convlab-2.

Thanks, I am not sure if DQN will work. Since in Convlab1, there are only 300 combination of actions in total, but in convlab2 it will be more than 8000. When using same dataset, did you know why the combination of Convlab2 is so much?

from convlab-2.

Hey, about the action space, do you gays have some clue? Like should I add some model of action encoder to Convlab2? So I can narrow the action space from 8000 to 300, by only considering the most frequent actions.

from convlab-2.

In ConvLab, we select the most frequent action combinations from all combinations in the dataset (https://github.com/ConvLab/ConvLab/blob/master/data/multiwoz/da_slot_cnt.json). You can try this approximation in ConvLab-2, too.

from convlab-2.

Please refer to #96

from convlab-2.