Closed rcoreilly closed 1 year ago
This is likely to be a more important issue in eboa -- the process of foveation is incremental as you rotate and it may take a few tries to foveate something -- by which time the novelty will be significantly decremented. It is much more robust to condition ACh / gating on whether you've already gated -- that is always reliable.
.LDTLayer.LDT.MaintInhib
= 2 works well (lower also works but takes longer at the start -- leaves more ACh around when maintaining. #SC Act.KNa.Slow.Max = 0.2
reduces the TrialSlow
novelty adaptation in SC to just take a bit of activity down as it stays around, instead of fully suppressing -- didn't do much to explore these params.
#SC Act.KNa.Slow.Max = 0.1
still suppresses eventually -- need .05? anyway, this is all working and much better.
maintaining a goal should inhibit SC -> ACh novelty signals. this may do more of the work vs. the complete shutoff accommodation novelty we use now. in particular, if you lose maint, ACh recovers, but the TrialSlow accommodation mechanism always shuts down novelty completely and that can lead to some hard-to-recover-from dynamics.