https://www.ncbi.nlm.nih.gov/pubmed/30122431
Introduction to the special issue on deep reinforcement learning: An editorial.
Author information
- 1
- Cognitive Science Department, Rensselaer Polytechnic Institute, 110 Eighth Street, Carnegie 302A, Troy, NY 12180, USA. Electronic address: dr.ron.sun@gmail.com.
- 2
- Google DeepMind, London, United Kingdom; University College London, United Kingdom. Electronic address: davidsilver@google.com.
- 3
- Thomas J. Watson Research Center, Yorktown Heights, NY, USA. Electronic address: gtesauro@us.ibm.com.
- 4
- School of Electrical and Electronic Engineering, Nanyang Technological University, Nanyang Avenue, Singapore 639798, Singapore. Electronic address: egbhuang@ntu.edu.sg.
- PMID:
- 30122431
- DOI:
- 10.1016/j.neunet.2018.08.001
However, there are still many questions and issues that need to be addressed with regard to deep reinforcement learning. For example, open questions with regard to deep reinforcement learning include:
How can one better extend reinforcement learning algorithms and systems to make them suitable for deep learning?
How does one make reinforcement learning (which are typically centered on values of states or state-action pairings) appropriately deep?
How does one do so without jeopardizing useful properties and performance characteristics of reinforcement learning?
What modifications and enhancements to learning algorithms can be made to accomplish deep reinforcement learning in an effective and/or efficient manner?
How can one make embedded knowledge within deep reinforcement learning systems explicit (i.e., generating explicit, symbolic, and directly usable knowledge) and enable metacognitive reflection and regulation?
How can deep learning help to facilitate planning?
How can hierarchical or modular approaches be applied to deep reinforcement learning?
How can one automatically generate intrinsic motivations and rewards as opposed to learning solely from extrinsic rewards from an environment? (This topic is discussed in more detail in the special issue.)
What theoretical/mathematical properties can be obtained with regard to deep reinforcement learning (e.g., convergence, stability, robustness, and optimality)?
How does one best apply deep reinforcement learning to real-world scenarios? (This topic is discussed in the special issue.)
DEEP RL 설계시 고려해야 할 사항들...
