I am trying to create a dialogue policy model on DSTC data. This model takes in a state of the conversation and outputs an act the machine must take. I am creating this model using reinforcement learning.
I am manually giving the rewards for each turn through long credit assignment with
-1 at each turn and at end,
(10 - turns) for success and
-2*(max_turns) for failure.
max_turns is kept at 20 turns for each conversation.
I have already trained over 1000 conversations yet the model does not seem to learn a bit. I wanted to know what common problems might be causing it to fail at all conversations. Am I training it wrong? Is my reward system incorrect? Or are 1000 conversations just not enough? Any help, tips that leads me in correct direction will be appreciated.