Justin Mason escreveu: >> $K3 = 1 was my first attempt but experiments showed that higher values >> produce better accuracy, with a maximum around 8. > > aha, I get it. That makes sense. Thanks for the explanation. > > So using K3=8 limits the maximum CF(F) to ~0.125? Doesn't that restrict > your probabilities to the [0.5-.125, 0.5+.125] range?
Even more: [0.5-0.125*0.5, 0.5+0.125*0.5] Is that desirable? The funny thing is that experiments say so, despite what may seem desirable to us. I guess there is room for research here. > I would have assumed we'd still want to allow really "strong" tokens > to reach nearly 0 or nearly 1.0. Yes, that's the normal first thought. But feel free to try other values for K1, K2 and K3. I didn't try other values after EDDC was published, not even after I changed the training method to TONE-HR. Maybe we'll get some surprise results. -- Fidelis Assis
