I recommend a feature that filters the training material by a minimum sensitivty between actions. Getting feedback on a 0 ev difference between actions is not productive, particularly when the trainer shows a strict prefernce for one action that is the same EV as another, especially when the spot is a triple split action.
The filter would make it so that the trainer would only draw from situations where the action is worse by a minumum amount of EV set by the user. ie .01bb .02bb .05bb etc