The approach I've taken, which I think is common and slow and easy to "fall off the bandwagon", is to write my solutions to my problems on my blog or email the J mailing list and see if the experts can improve on my code. They always can, if they care enough to show me, and they usually do. But as I mentioned in my other comment, I think one of the wonders of APL/J/etc. is that it seems to share even less with conventional programming languages than it seems to at first because the way APL programmers decompose problems is radically different. I often find myself trolling through the dictionary looking for things that I wouldn't need if I didn't break the program down into as small pieces. Choosing the right verb is partly tactical, but the whole strategy for attacking problems is so unconventional you are often looking from the wrong vantage point.
I don't know if there is a royal road to APL that addresses this, but I share Hillel's desire for one. I suspect there isn't, and the only way to really get better is: read lots of code, write lots of code, get feedback on your code, repeat.