what is usage of data/english/dialogueN.txt
Closed this issue · 4 comments
yoosan commented
Hi, I have a question about the usage of dialogues.
The dialogues in data/language file is just used to train a classifier?
cuayahuitl commented
Hi Yao,
the unique words in those dialogues are used as features in the system. In
addition, those dialogues are used to train a statistical classifier that
SimpleDS uses to (1) select the most likely actions given word-based
features, and (2) provide data-like rewards. Notice that during training
the system will potentially induce its behaviour from a large number of
different dialogues, derived from combining different system responses with
semi-random user responses.
I hope this answers your question.
Best,
Heriberto
yoosan commented
Thanks Heriberto,
I got your meaning since i have read your paper SimpleDS
. The dialog examples in data/english
is used to compute the rewards, and to narrow action space depending on probability threshold? We know that deep Q-learining with experience replay is well suited for unsupervised learning, there is no need for preparing supervised data. These 6 dialogs are not the training examples in RL.
wxtyixiao commented
cuayahuitl commented
Hello Xiaoting,
I don't know if I have already answered your question, but please set the
number of dialogues to a larger number. One dialogue is for testing
purposes, and the trained policy (according to the config.txt file) is
saved every 2500 dialogues. Hope it helps.
Best regards,
Heriberto
…On Thu, 27 Sep 2018 at 10:26, Xiaoting ***@***.***> wrote:
Hi, I have a question about the running procedure of runclient.js.
During the training, the policy and the output hadn't been saved in
results. Is it right?
[image: _20180927172202]
<https://user-images.githubusercontent.com/24651880/46136543-70426c00-c27a-11e8-9887-8133b6d72463.png>
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#3 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AO5H1OS-woKSdU8sme2d3F-3lvlCNL_Eks5ufJnLgaJpZM4Jpd4k>
.