Learning Expected Emphatic Traces for Deep RL