[Home]David Alves/JALT

Robo Home | David Alves | Changes | Preferences | AllPages

Just A Little Test

This is the bot that I test new ideas in.

Version History

JALT 2.x Discussion

Hmmm - I wonder, is the gun still unsegmented? You seem to get a better hit rate on RMX than anything I've tested bar a few pattern matchers, DT and my own gun. Maybe that could explain a few things. :-) -- Jamougha

No, the gun is not unsegmented. Sorry, I was unclear about that. The self-gun that the enemy fires at JALT with (for adaptive movement) was unsegmented, not the gun that JALT fires with. In 2.02 the self-gun is no longer unsegmented, it's segmented by distance into 4 buckets, but on an 800x600 field it probably won't make much difference, since only 0-300 and 300-600 are going to be used much. I'm not convinced that highly segmented self-guns are useful. By the way, which version is it that gets the high hitrate vs. RaikoMX, 2.01 or 2.02? It'd be pretty funny if it was 2.01, watch it fight Wolverine... :-P --David Alves

Ah, should have realized that! It's 2.02 which got the high hit rate. I'm fairly snappy about checking new versions out. ;-)

I'm also not convinced that highly segmented self-graphers are worthwhile... except where they happen to just coincide with the opponent's gun. Figuring out when that is the case, without resorting to 40-50 if(e.getName() ==...) statements, is something else. :-) -- Jamougha

Every time the opponent fires eliminate all the guns which are telling you to use a different guess factor from what the opponent used; you'll quickly eliminate all of your guns or all but one. You'll probably need to use something like AutomatedSegmentation to actually generate the exact gun of your opponent, though. -- Kuuran

'Fraid it's not so easy. :-\ Even if the gun you're using is absolutely 100% compatible with theirs, you still don't get hit where you would expect to get hit very often (~25%)... the simple reason being that AM-bots don't hang around where they expect to get hit much. :p If you watch RMX against FloodMini then it may not get hit in the first couple of rounds, but eventually discrepancies creep in. I have a feeling, based on tests so far, that virtual enemy guns will suffer from the same problem as virtual targeting; too little difference between the data with too much noise over too few battles. Except with even less data. -- Jamougha

Are you sure you aren't just experiencing some simulation error? Are you working with the same timeline ABC wrote up elsewhere on the wiki (EnemyWaves I think?), or are you firing into a different segment from your opponent? If you're using their exact gun you really shouldn't be getting hit at all (when clean of walls) unless there's some discrepency you haven't accounted for, and the only one that isn't easily controlled would be skipped turns... As for too little difference between data with too much noise over too few rounds, I don't really see how it would suffer from this any more than the waves your enemy fires at you if you follow their procedure exactly. You're taking the same data and analyzing it the same way, any error introduced should be identical and affect your choice of GF to shoot at in the same way. -- Kuuran

Unfortunately opponents do all sorts of things to mess you up; skip turns, loose radar locks, fail to acquire you at the same time you acquire them, assume a different bullet power to you because they entered a different distance segment or expended all their energy, fire when their gun is pointing a few degrees from where it should be, have buggy guns, etc. etc. Bear in mind that if your movement is flat - and I mean really flat - it takes only the slightest discrepancy between their gun and your simulation to make tham fire somewhere completely different to where you expected. They basically end up using random targeting, which pretty well reflects the hit-rate which FloodMini-based guns get on my movement (plus a little bit because my simulation of my own movement sucks :-) ). It would be amazing if anyone were to push bullet dodging past the 'random targeting limit'. -- Jamougha

I was surprised that you said JALT was getting a good hit rate vs. RaikoMX so I ran a TC. It scores 82.5%. Are you sure your test was thorough enough to get a complete picture of JALT's accuracy? Sounds to me like you didn't run enough battles. Then again, performance in 35 rounds when JALT is allowed to pick the bullet power and range is very differerent from performance in 500 rounds where it has to fire power 3 bullets and can't change its range, so a TC probably isn't very indicative. :-p Maybe I'll try a fast learning TC. --David Alves

This is over two 250 round battles. It scored fractionally under 10% of the time, which is definitely unusually high. Looks like it's either specifically good at picking up on AM, or it's buggy and my movement couldn't figure out what it was doing. ;-) -- Jamougha

It scored very low in the fast learning TC too, again around 82%. I think I've taken most of the bugs out of it. Perhaps your movement and my targeting are just segmenting on different things? Try 2.1, it does *much* better in both TCs. ;-) --David Alves

I'm now the proud owner of the highest-ranked US bot. :-D --David Alves

Do you want to tell something about this new targeting ideas? -- rozu

More intelligent segmenting, and a couple of tweaks to improve accuracy at very close range. 2.11 is up now, maybe I can break 1900 by the end of the day. :-) --David Alves

Welcome back to top-10! -- PEZ

Thanks, PEZ. :-) --David Alves

/JALT 1 Discussion

Robo Home | David Alves | Changes | Preferences | AllPages
Edit text of this page | View other revisions
Last edited March 19, 2004 20:56 EST by David Alves (diff)