search

lifelong-learning-systems / rlblocks

Reinforcement Learning Blocks for Researchers
MIT License
0 stars 0 forks source link

adding scp loss to torchblocks #24

Closed neilfendley closed 2 years ago

neilfendley commented 2 years ago

Adding in the sliced cramer preservation loss. Currently setting the parameters to 10 projections and lambda = 1.

neilfendley commented 2 years ago

Have updated so now it can merge with base

coreylowman commented 2 years ago

I think this looks really good! 🎉 🚀 I bet HRL will like this implementation.

lets merge after @ginoperrotta reviews

ginoperrotta commented 2 years ago

I created #25 to fix non-repeatability in this (and in all of our DQN agents). That has helped clarify that the SCP agent is equally capable at learning the lava environment; the review comment was based on a particularly poor (but not repeatable) performance.