#deeprl — Public Fediverse posts

This #DeepRL paper from University of Alberta seems quite cool:

"Deep reinforcement learning without experience replay, target networks, or batch updates"

As the title says, they succeeded in training deep RL networks in streaming setting getting rid of replay buffers.
The main tricks for that to work seem to be signal normalization and bounding the step-size 🤯

💻Code: http://github.com/mohmdelsayed/streaming-drl
📄Paper: https://openreview.net/pdf?id=yqQJGTDGXN

#AI #RL #DeepLearning

#AI #MSc_AI #Python #DeepLearning #AIEthics
#computationalcognition #creativeindustries #deeprl #ReinforcementLearning

MSc in AI. Registration is open.

Requirements:
First (or, pending on CV, upper second class) BSc/BEng/BA in computer science, mathematics, physics, computer engineering, psychology or biology.
Competence in Python and Mathematics

Interested?
For more information and to download a booklet, please visit https://cit-ai.net/CitAI-MSc.html

To Apply:
https://www.city.ac.uk/prospective-students/courses/postgraduate/artificial-intelligence/2023#accordion513943-header513943

Last year, a French #ai Athénan won 11/23 of the games at the 24th ICGA Computer Olympiad. Nice.

Even nicer: it ran on a single GPU, yet it challenged cutting-edge #deepRL-based AIs requiring hundreds of GPUs!

Here is the paper about its algorithm (a minimax): https://arxiv.org/pdf/2012.10700.pdf

👉 New Preprint available #arXiv
http://arxiv.org/abs/2205.09738
AIGenC is a model for creative problem-solving in a deep reinforcement learning agent as a step forward to solving the problem of generalisation in #AI #deepRL