search

ramack / ActivityDiary

Android diary for any kind of activities
GNU General Public License v3.0
73 stars 37 forks source link

Use ML techniques for conditions #182

Open ramack opened 6 years ago

ramack commented 6 years ago

Currently the different conditions are evaluated and theirs output just summed up for all activities.

This sum should be replaced by a neural network with an input node per condition and an output node per activity to calculate the current likelihood for each.

Or one network per activity?

=> a design to be done first...

ramack commented 6 years ago

As we might have dependent cobditions e.g. breakfast is later on sundays than on workdays we qould newd a very broad NN or conplex neurons. Maybe a multiplication layer is helpful. With a pairwise pre initialization but also some randomly initialized ones. SpeIal backprop needs to be done here.... Also we shoul every now and then derate the not current activities, while rating up the current one. On manual selection the last one shall be derated mire and the new one even more rated up...