Sequence generative adversarial networks for music generation with maximum entropy reinforcement learning

Mathias Bjare, David Meredith

Publikation: Bidrag til bog/antologi/rapport/konference proceedingKonferenceartikel i proceedingForskningpeer review

Abstrakt

We introduce using maximum entropy reinforcement learning (MERL) for training sequence Generative Adversarial Networks as a technique for obtaining more diverse samples and alleviate mode drop (MD) and mode collapse (MC) problems. We implemented generators using the ordinary REINFORCE algorithm, REINFORCE with reward baseline and MERL REINFORCE. We trained the models to learn to generate music using the Nottingham data set. We observed that, without pre-training, the algorithms fail to produce high reward trajectories. We showed that the pre-trained REINFORCE models mainly explore trajectories of their initial policies and we argue that the method might be more suitable for fine-tuning models than learning generative distributions from scratch.
OriginalsprogEngelsk
TitelProceedings of The 2020 Joint Conference on AI Music Creativity (CSMC + MuMe 2020)
RedaktørerBob Sturm
Antal sider9
UdgivelsesstedStockholm, Sweden
ForlagKTH Royal Institute of Technology
Publikationsdato2020
ISBN (Trykt)978-91-519-5560-5
StatusUdgivet - 2020
BegivenhedThe 2020 Joint Conference on AI Music Creativity - Royal Institute of Technology (KTH), Stockholm, Sverige
Varighed: 19 okt. 202023 okt. 2020
https://boblsturm.github.io/aimusic2020/

Konference

KonferenceThe 2020 Joint Conference on AI Music Creativity
LokationRoyal Institute of Technology (KTH)
LandSverige
ByStockholm
Periode19/10/202023/10/2020
Internetadresse

Fingeraftryk Dyk ned i forskningsemnerne om 'Sequence generative adversarial networks for music generation with maximum entropy reinforcement learning'. Sammen danner de et unikt fingeraftryk.

Citationsformater