search

xbpeng / mcp

Implementation of multiplicative compositional policies.
10 stars 1 forks source link

Variance linear output layer? #2

Closed AlexanderKoch-Koch closed 4 years ago

AlexanderKoch-Koch commented 4 years ago

Hi, in the paper it says: "a linear output layer that produces μi(s) and Σi (s) for each primitive". However, in the formula to calculate the resulting distribution you have to divide by the variance. So a zero variance would be a problem. And I don't know if a negative variance makes sense?

xbpeng commented 4 years ago

Right, the linear output layer actually outputs the log variance. We apply an exp to prevent non positive variance.