Hi,
I'm trying to play around with the parameters of the DGMG, but it seems that I get messy figures or alternativly too abstract (usually straight lines), but I can't hit the middle.
Are there maybe other parameters I should change?
I'm using Adam optimizer, 30 epochs and loss function is defined like this:
loss = - model(actions = actions).exp()
(BTW, how can I add images here...? )
you should not use
- model(actions = actions).exp()
as the loss, just
model(actions = actions)
(without the .exp() )
it is numerically more stable to increase the log(prob) vs. the prob itself.