Today’s RL algorithms are great at exploiting a particular environment, terrible at using that knowledge in new situations. Here’s a new environment which is already helping us understand why, and which may help develop RL algorithms that generalize:
OpenAI
 @OpenAI
6 Dec 2018
We’re releasing CoinRun, an environment generator that provides a metric for an agent’s ability to generalize across new environments - blog.openai.com/quantifying-…

Dec 6, 2018 · 4:44 PM UTC

8
64
1
196
This is our third major attempt in the past two years (Universe, Retro Contest) to develop a platform for RL generalization. Each time, we’ve made the task easier — but more focused on the core generalization challenge. Already seeing promising results on CoinRun.
3
7
42
Replying to @gdb
Care to share information on the compute resources used to evaluate this?
Replying to @gdb
FYI, this looks extremely similar to a workshop paper being presented at #NeurIPS tomorrow and I don't see any reference to it.. arxiv.org/pdf/1806.10729.pdf @OpenAI @gdb
1
2
Replying to @gdb
Finally!
Replying to @gdb
Very insightful paper. Especially the part showing that classic regularization techniques improve generalization in RL. However, where CoinRun seems to have fixed rules, GVG-AI offers a large suite of diverse and customizable environments, which I find more useful.
3
9
Replying to @gdb
Alright Greg I’ve been following you for awhile just transferred from finance to a software engineer give me a study course I want I be working with you and the crew in 2-5 years I understand there’s a lot of math and nodes but apparently I’m a nerd now
Replying to @gdb
Amazing work! Do you think that “old” metrics for general intelligence like this one (Legg and Hutter, 2007) can also be practical for modern RL research? (K is Kolmogorov complexity of environment)
Replying to @gdb
Has the team considered entering the @gvgai competition? It's been running for a few years with a whole range of games with procedurally generated levels, including a learning (model-free) track for RL algorithms.
3
6