The performance of dialogue system based on deep reinforcement learning (DRL) highly depends on the selected hyperparameters in DRL algorithms. Traditionally, Gaussian process (GP) provides a probabilistic approach to Bayesian optimization for sequential search which is beneficial to select optimal hyperparameter. However, GP suffers from the expanding computation when the dimension of hyperparameters and the number of search points are increased. This paper presents a meta learning approach to carry out multifidelity Bayesian optimization where a two-level recurrent neural network (RNN) is developed for sequential learning and optimization. The search space is explored via the first-level RNN with cheap and low fidelity over a global region of hyperparameters. The optimization is then exploited and leveraged by the second-level RNN with a high fidelity on the successively small regions. The experiments on the hyperparameter optimization for dialogue system based on the deep Q network show the effectiveness and efficiency by using the proposed multifidelity Bayesian optimization.
|頁（從 - 到）||839-843|
|期刊||Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH|
|出版狀態||Published - 2019|
|事件||20th Annual Conference of the International Speech Communication Association: Crossroads of Speech and Language, INTERSPEECH 2019 - Graz, Austria|
持續時間: 15 9月 2019 → 19 9月 2019