Search

20 Feb 2018: On-line active reward learning for policy op-timisation in spoken dialogue systems.

20 Feb 2018: Recent workby Graves et al. (2014) has demonstrated that anNN structure augmented with a carefully designedmemory block and differentiable read/write op-erations can learn to mimic computer programs.Moreover, the

20 Feb 2018: By op-timising directly against the desired objective func-tion such as BLEU score (Auli and Gao, 2014) orWord Error Rate (Kuo et al., 2002), the model canexplore its output space

20 Feb 2018: Hence defining an op-timal summary policy is not so obvious. If f is chosenwell, however, then one could hope that the optimal ac-tion is dependent only on f (b).

20 Feb 2018: Note that the reward model and the dialogue policy are being jointly op-timised during the sequence of dialogues.

20 Feb 2018: This Gaussian process op-erates on a continuous space dialogue rep-resentation generated in an unsupervisedfashion using a recurrent neural networkencoder-decoder.

20 Feb 2018: increases. Pop op-erations are then performed where possible, the tree is prunedand identical nodes are joined so that the number stays constantor decreases. ... Error bars indicate 99% con-fidence intervals. This demonstrates the competitiveness of the

20 Feb 2018: A comparison between the three op-tions is included in the experimental evaluation. ... whilst suffering initially.We hypothesise that the optimised SL pre-trainedparameters distributed very differently to the op-timal A2C ER parameters.

20 Feb 2018: In Section 3, the grid-based ap-. proach to policy optimisation is introduced followedby a presentation of the k-nn Monte-Carlo policy op-timization in Section 4, along with an ... 5 ConclusionIn this paper, an extension to a grid-based policy

20 Feb 2018: The op-tions for each random decision point are reason-able in the context in which it is encountered, buta uniform distribution of outcomes might not re-flect real user behaviour. ... Many of the decisions involvedare deterministic, allowing only one