Yes, the Alg Desc has only a single parameter for the subsumption threshold, and in my Oblique paper I use two different ones. So you should use different parameters in the procedure COULD SUBSUME, according to which kind of subsumption is calling it. However, the same epsilon_0 is used in both.
Theta_assub = 1000 in order to make sure that the potential subsuming classifier is actually accurate. The WBC data set is large, and it takes a long time to be sure a classifier has seen all of it enough times.
For some reason, longer thresholds are more important for action set subsumption than for GA subsumption. It's not clear to me why. That is, in the WBC experiment, making theta_GAsub greater than 50 didn't make much difference, but it made a great deal of difference to increase theta_assub.
The threshhold for GA subsumption is 50 and for Action set is 1000, right? If so, the method COULD SUBSUME (in the Algorithmic Description paper, Springer-Verlag version) will be different for them, but the epsilon_0 is 1 for both of them. Why 1000 is used?