BedMaker
The highlighted comment was created in this revision.
Glad to see someone else enthusiastic about BedMaker. =) I also dropped Darkcanuck an email a few days ago. I'm not sure what's up - he's usually pretty responsive. Maybe a summer vacation?
Anyway, while it's not a good permanent solution, let me know if you want me to run some configurations for you and post the RoboRumble challenge files, to get you started. For instance, my current test bed is like: "400 bots that Diamond scores 60-92 against" (~top half of the rumble, bedmaker.pl -n 400 -r roborumble -ref voidious.Diamond -min 60 -max 92), and I think my next one is going to be "500 bots that Diamond scores 55 - 99 against" (ie, large random sampling of most of the rumble, excluding DrussGT and bots I score almost perfectly against).
Thanks!
I'm sure it's vacation and summer stuff. No biggie. I'm really looking forward to making some test beds so that I can try out some fixes and tweaks, give them an evening/night run on my computer and see how they fare without bothering the Rumble with them.
It's been confounding to me how often I'll make a change, test and see improvement against 8/10 bots, and see that alteration tank in the Rumble! Not to mention the time taken in manual testing. Much more fun to just code, and see the results in the morning locally.
I've been trying to get up near CassiusClay, YersiniaPestis et. al. in Rumble score before adding on more advanced stuff like Bullet Shadows and Gun Heat Waves (both of which I've got my code in a state to pretty easily tack on, thankfully!) It's been a heck of a struggle though, and I'm fighting for every 0.1 APS it seems. Good ol' law of diminishing returns.
If you wouldn't mind generating a challenge file for me, I'd be appreciative! How about....
"150 bots that deBroglie rev0108 scores 57-95 against"
That should get me a decent size test bed of bots that I'd like to see myself do much better against. There hasn't been too much churn version-to-version above the 95 percent line.. and there's been lots of churn below ~57, most of which will take more advanced work to really beat anyway.
Feel free to tell me if you think that's a poor choice for an 'Am I helping or hurting this bot with this change' type of test bed, or if it should be bigger or smaller.
Sure, here you go - I made a second with the same parameters in case you hit errors with any of the bots, or they're super slow, and you want to swap some out by hand: [1] [2] One nice thing bedmaker does is copy just those bots into roboresearch/robocode_bots, so I guess you'll have to just copy your whole rumble/robots dir in there unless you want to write a script or something. =)
That seems like a good cross-section to test general rumble performance. I hate wasting CPU cycles re-running battles with almost no variance, like bots I score 99+ against, but at the same time you have to be careful not to specialize too much on the bots you have lots of room for improvement against. I certainly haven't mastered the art of choosing test beds, but big test beds across a large section of the rumble seem pretty safe. They'll at least show you pretty quickly when you've really screwed up. =)
Lately I've been running ~2000 battles, which seems accurate to within 0.05 or 0.1. That's made easier now that I can do that in 2-2.5 hours on my new machine, though - previously it was taking me 2+ hours for one season of a 250-bot test bed!
So you'll have to feel it out for yourself as far as your bot and what changes you're making (and how patient you are =)). Like with some changes, I just want to know it's not totally screwing things up before testing it in the rumble. While tweaking parameters little by little, I have to run a lot more battles to know what's going on.
Excellent! After culling a few bots that were throwing BufferedInputStream errors (seems like it was bots that like to output to the external Robocode terminal instead of their own System.out) It's cranking away.
Is there a convenient way to view RoboResearch output for test beds this big other than copying to wiki? That results window is awfully crowded. :)
Thanks so much for the help, Voidious!
No problem! I don't know, my test beds are usually either so big that I only check the overall score, or so small (like my 5 worst matchups) that I can just remember the order. =) RoboResearch also has a GUI which probably has nicer output, but IIRC it puts all the scores on one line, which might be a few thousand pixels wide with 150 bots.
Oh, "results window", maybe you are using the GUI? I use the command line ("CLI"), which I like, but it seems most people do prefer the GUI.
