Cookies help us deliver our services. By using our services, you agree to our use of cookies. Learn more

crowdAI is shutting down - please read our blog post for more information

NIPS 2017: Learning to Run

Reinforcement learning environments with musculoskeletal models


Competition rules query

Posted by dkudenko over 1 year ago


In the competition rules it states that “You are not allowed to use external datasets (e.g., kinematics of people walking),”. Does this also mean that any form of domain knowledge (e.g. in the form of heuristics) is not permitted, i.e. that the solutions should be a tabula-raza approach to RL?




Posted by Yongjin  over 1 year ago |  Quote

I have a similar question. In the previous posting, we were allowed to use several consecutive observations to estimate or compute the speed and the acceleration of the head. I wonder to what extent we can modify or manipulate the observations. For example, can we normalize the values of observations or tweak them manually using some heuristics as long as we do not use some physical laws for kinematics of people walking? However, for computing an acceleration, we are actually using the laws of physics although it is very basic.

Posted by spMohanty  over 1 year ago |  Quote

Hi @dkudenko, @Yongjin,

The idea of the challenge is to explore the feasibility of approaches like Deep Reinforcement Learning for a classic problem from Musculoskeletal modelling. And ofcourse, minor adaptations to the observations are indeed allowed, as long as the transformations you use on the observations are not over engineered for the particular problem.