Szczegóły publikacji
Opis bibliograficzny
Towards understanding of deep reinforcement learning agents used in cloud resource management / Andrzej Małota, Paweł Koperek, Włodzimierz FUNIKA // W: Computational Science – ICCS 2023 : 23rd international conference : Prague, Czech Republic, July 3–5, 2023 : proceedings, Pt. 2 / eds. Jiří Mikyška [et al.]. — Cham, Switzerland : Springer, cop. 2023. — (Lecture Notes in Computer Science ; ISSN 0302-9743 ; LNCS 14074). — ISBN: 978-3-031-36020-6; e-ISBN: 978-3-031-36021-3. — S. 561–575. — Bibliogr., Abstr. — Publikacja dostępna online od: 2023-06-26
Autorzy (3)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 147736 |
|---|---|
| Data dodania do BaDAP | 2023-07-20 |
| DOI | 10.1007/978-3-031-36021-3_55 |
| Rok publikacji | 2023 |
| Typ publikacji | materiały konferencyjne (aut.) |
| Otwarty dostęp |  |
| Wydawca | Springer |
| Konferencja | 23rd International Conference on Computational Science |
| Czasopismo/seria | Lecture Notes in Computer Science |
Abstract
Cloud computing resource management is a critical component of the modern cloud computing platforms, aimed to manage computing resources for a given application by minimizing the cost of the infrastructure while maintaining a Quality-of-Service (QoS) conditions. This task is usually solved using rule-based policies. Due to their limitations more complex solutions, such as Deep Reinforcement Learning (DRL) agents are being researched. Unfortunately, deploying such agents in a production environment can be seen as risky because of the lack of transparency of DRL decision-making policies. There is no way to know why a certain decision is made. To foster the trust in DRL generated policies it is important to provide means of explaining why certain decisions were made given a specific input. In this paper we present a tool applying the Integrated Gradients (IG) method to Deep Neural Networks used by DRL algorithms. This allowed to obtain feature attributions that show the magnitude and direction of each feature’s influence on the agent’s decision. We verify the viability of the proposed solution by applying it to a number of sample use cases with different DRL agents.