Model-based Distribution Agents for Randomly Delayed Reinforcement Learning
Dnr:
NAISS 2025/22-675
Type:
NAISS Small Compute
Principal Investigator:
John Wikman
Affiliation:
Kungliga Tekniska högskolan
Start Date:
2025-05-12
End Date:
2026-06-01
Primary Classification:
10210: Artificial Intelligence
Webpage:
Allocation
- Alvis at C3SE: 1000 GPU-h/month
- Mimer at C3SE: 500 GiB
- Klemming at PDC: 500 GiB
- Dardel-GPU at PDC: 200 GPU-h/month
- Dardel at PDC: 1 x 1000 core-h/month
Abstract
We are developing a new kind of algorithm where agents in reinforcement learning can adapt to random unobservable delays. This is done by training up a model of the system and using that to compute distributions over possible future states to make more informed decisions. We use deep neural network to represent the models.