jharwell / sierra

Automation framework for the scientific method in AI research
MIT License
18 stars 1 forks source link

bugfix/105-karp-flatt-projection-incorrect #105

Closed jharwell closed 5 years ago

jharwell commented 5 years ago

It always does projective performance for expN relative to exp1, instead of exp(N-1), which is much more accurate, in my opinion, esp. with respect to the GP-DPO approach which has terrible performance with just 1 robot.