Bram Grooten
@bramgrooten.bsky.social
PhD student at the TU Eindhoven. Researching deep learning, RL, sparse neural networks, robotics.
Haha true, the last part was mentally the toughest!
February 4, 2025 at 1:44 PM
Haha true, the last part was mentally the toughest!
Haha thanks, no worries. It was kind of spontaneous. I started the day with 30km cycling, then 1km swimming. Then I realized I could make this into a quarter triathlon, so ran 11km and added an extra 20km on the bike!
February 4, 2025 at 7:10 AM
Haha thanks, no worries. It was kind of spontaneous. I started the day with 30km cycling, then 1km swimming. Then I realized I could make this into a quarter triathlon, so ran 11km and added an extra 20km on the bike!
The swimming part was by far the hardest for me. And about the cycling: I was on my electric bike, so let's call it an electric quarter triathlon
February 3, 2025 at 8:42 PM
The swimming part was by far the hardest for me. And about the cycling: I was on my electric bike, so let's call it an electric quarter triathlon
I'm also curious to hear @shibhansh.bsky.social's opinion. On what timescale did loss of plasticity usually occur in your RL experiments? :)
January 10, 2025 at 4:54 PM
I'm also curious to hear @shibhansh.bsky.social's opinion. On what timescale did loss of plasticity usually occur in your RL experiments? :)
Thanks! That's interesting. Would indeed make sense that CL issues happen more often in more complex environments that require longer training.
January 10, 2025 at 4:51 PM
Thanks! That's interesting. Would indeed make sense that CL issues happen more often in more complex environments that require longer training.
How long does an RL agent usually need to be trained before you notice CL problems, would you say?
January 10, 2025 at 6:48 AM
How long does an RL agent usually need to be trained before you notice CL problems, would you say?
What I often see is that people just take buffer size = time steps. Up to a max of about 2M
December 11, 2024 at 2:36 PM
What I often see is that people just take buffer size = time steps. Up to a max of about 2M
We worked on making PbRL more robust to noisy envs recently:
arxiv.org/abs/2406.06495
arxiv.org/abs/2406.06495
Boosting Robustness in Preference-Based Reinforcement Learning with Dynamic Sparsity
For autonomous agents to successfully integrate into human-centered environments, agents should be able to learn from and adapt to humans in their native settings. Preference-based reinforcement learn...
arxiv.org
November 22, 2024 at 9:13 AM
We worked on making PbRL more robust to noisy envs recently:
arxiv.org/abs/2406.06495
arxiv.org/abs/2406.06495