Numbat - Wompi/robocode-bots GitHub Wiki

The development diary of Numbat. The official version history can be found at RoboWiki (Numbat). The diary is just to keep track of the thoughts and odds while programming on this bot. I guess it is not appropriate for the official wiki and therefore i link it to this repository.

Current Notes: Things to try

Date: 10/08/2012 Size v1.8:

The gun and radar got some minor improvements this version. The gun collects pattern length on 3 now and the target velocity is collected on 16 segments now. The radar got a new rule which improves the awesome factor a little - if the GunHeat is near shooting the radar will no longer search for targets. This makes the radar working very smooth - scan all targets - lock the mainTarget - wait till the radar wave reaches the bot - if gunHeat is high enough sweep the radar to update the other targets otherwise hold the lock till shooting and then go for search.
While working on the gun i figured out that i have to make a massive change to the gun to make it work against oscillators and especially Tron. These bots can not be hit because the use of the 'most visited pattern' is not working on this kind of movement (its always 50:50). I also should consider to change the target state to real angles to make the code smaller. Well we will see

Credit: nothing new
...off we go

Date: 08/08/2012 Size v1.7:

Hmm my new formatter messed up the comments but for now i'm to lazy to fix this. Anyway i try to switch to github and comment the code while punching it in. After aeons of testing, searching and crying i finally found a huge bug within my gun and from now on i hit wall with an impressive hitratio. The bug was very well hidden and i stepped over it by accident. Well at least the gun is now a little better and i started to tweak some combat settings. Maybe i can go back to dynamic pattern and zero pattern we will see. I put all my botcode on github and i'm quite happy about this. I changed a couple of minor things too but have forgotten was it was.

Credit: hmm
... yehaaa a new version

Date: lost Size v1.6:

Hmm the summer is killing my brain and i can't think straight anymore. I tested lots of stuff with the gun but it brought me nothing. I decided to switch back to a cleaner code base, mostly because the 'hidden' skipped turns drive me crazy right now. I changed just minor stuff to be more precise in the gun, got rid of the dynamic key length, brought back the zero patterns and some other stuff. What concerns me most is that i cannot hit Walls with a hit ratio above 90 - there is something really wrong with my gun :(. I have to admit that got a little lost with my code and maybe should take a break until the weather cools down a little. All debug and paint things are disabled and commented in this version

Credit: hmm
... toooo hot

Date: lost Size v1.5:

After a very long and frustrating time of debugging and refactoring i have hopefully found something that is worth a release. I have to admit, that this version is a very premature release and is not fully tweaked and reviewed. The reason behind this - i spend the whole day to find something that could increase my APS and right before i would give up it came to me. My test bets get busted with this version and i'm right now very exited to see how it works in the rumble. Right now i ask myself, why are always the obviously stupidest ideas the ones that bring most success. Well lets see what i have changed. First and most visible change is the code base. It got an complete overhaul and is now more OOP (still don't like it) and is highly modularized. From now on i can make new guns, radars, moves, fire strategies and what not and it is instantly ready to use. Maybe in future versions i switch guns and radars to fit the current battle situation better like i do now but only for maintenance reasons. The radar is completely changed and finally deserve the name code. The function is still the same but with way less code and more stability. I really like it now and it works superior (by my standards). It is very embarrassing if i see how stupid the old 'code' was and now it has the same function and is such a tiny piece of code. Well it works good :) Hmm what other stuff is new.... hmmm? Ah yap. The debug properties is now something that i have fallen in love with. From now on i can just tap some keys and can switch of and on the related debug informations. The informations will increase in future versions but the framework is ready and works. (r = radar, g = gun, s = misc, m = move) (right now everything is enabled). The graphics are very basic and i disabled most of them. Only radar and some target informations will be shown. (at least it shows how the radar works). Last, but not least, the main change and the one that i hope to get some APS for is the gun change. I spend a lot of time to research the gun behavior, tried to play with a dozen different state keys, reinforced pattern change (maybe the wrong term, i just tried to revisit the used pattern and would adjust them) This worked not very well but it looks like it could work somehow (maybe bearing offsets or something). In the end i was so frustrated that i just made an 'rage' change and the all used velocity to the mean velocity average. You will guess it - jep it worked not very well. But it showed that the influence of the velocity on the pattern has some correlation to the hit rate. A couple of cigarettes later i was pretty sure that a short time average might work better than the current velocity. First try was 10 ... to much 5 better but lets try 3 ... bingo! I'm sure it will not bring that much with just this little try. But the direction is clear. Hmm i guess it is enough text ...

Credit: as usual
... lets go

Date: lost Size v1.4: 7400+

Awesome ranking at 11 general melee and the crowd goes wild :). The gun starts to kick in and supported by the overall synergies it performs better than Wallaby. This version has some minor tweaks to the radar. Look for the best angle to scan and find faster at start, and stop start scan if all opponents found. The gun is now really good at memory management because of the dynamic match key ability. This brings the saved pattern from almost 1.000.000+ to just 100k per melee battle. The synergy is if the enemies have long pattern (up to 120 i have seen) they have not much pattern to save, the other side is a strong movement with pattern around 5 and they have also not much pattern because of the small key. Pretty impressive right :). The gun holds his hit ratio and my gut tells me it could even be better with the dynamic match key. I changed the fire power selection to something i have read on the RoboWiki. Paul Evans stated there it might be better to shoot just 1.75 bullets because the net gain might be the same by increasing the overall survivabillity. You know what, i can say it magically works. I got the same score against all my test bets by just shooting 1.75 bullets most of the time. This guy is probably from another star or was raised by magical elves :). Anyway he deserves the credit and i i'm glad i can give it. Lets see how it works in the rumble. The debug got some shiny new property printout and can be seen in the debug property tab (key 'd')

Credit: Paul Evans by finding a magical number
... lets go

Date: lost Size v1.3: 6000+

Nice APS jump in the last version and hopefully this one will do also good. The gun got some minor tweaks, like saving patterns from just scanned targets (not only main target). But most attention got the movement. It is still just a melee movement but i hope the little changes made it competitive at least till top 10. The start movement is free and therefore no more wall hits at start and searching a good position to start the rumble. This also increased the near bot behavior and should be way more agile then the last one. The radar got a little workaround for the bug (still looking for it :(). I also fixed a couple of bugs which i can't remember anymore (mostly stupidity bugs like wrong delta calculations and such stuff) All in all i start to like this bot because it shows some potential and i haven't started to improve the gun by now. The hit rate against nano/samples and other weak mover is quite impressive. The graphic is now enabled (press 'p') and the code goes up and up :) Next step should be watching the gun a little closer, especially the memory use is something i haven't put interest after the last changes. But i hope it is ok and should be around 200MB per melee battle.

Credit: Numbat at RoboWiki
... lets go

Date: lost Size v1.2: 5000+

I decided not to hold me back in code size any longer and therefore i refactored the code to something more modular. The positive side effect is, that i don't get lost in the code anymore. The gun is still Single Tick Pattern but i enhanced the usability a little. The radar is still rude and i don't like it but for now it is enough (the radar has still a major bug where the bot ends as siting duck). The target holds now way more informations, like start pattern length (correlates to HitRatio) and some other goodies. Finally i managed to bring everything ready to work within run (needs still lots of work to be perfect). The new score targeting rule worked very good in testings. The target will now not be changed if i collected a lot of score and would waste the bonus if i switch to another target. Anyway, still lots of stuff to do and i don't think it will perform better then the last version but i give it a go.

Credit: Numbat at RoboWiki
... lets go

Date: lost Size v1.1: 1824

The first release was not that bad and hopefully this one will advance a little bit further. I cleaned the code and it is now almost mini class size. The gun got after a little research his first improvement and can now switch the pattern if the target would hit the wall. The guessed pattern takes in this case the most used next step and calculates the pattern based on this step till the bullet reaches the target. I also changed the combat variables to be more cowardly and the radar locks even more now. The radar got no change and is still bugging me. The data structure for the patterns works now very smooth and can very quick give the related informations. Also are no longer pattern for size zero saved, which saved almost 400k entries per match. The match key is decreased to 30 because it turned out that this quite a good match for almost every bot. The gun is now very fast in adapting movement changes. The first test have shown that the performance against top 20 bots is worst than ever. But everything over top 20 should be hit quite well, especially almost every nano bot is now a good pray.

Credit: look at Numbat at RoboWiki
... lets go

Date: lost Size v1.0: 1864

New bot and new hopes :). This bot utilize a SingleTick pattern matcher (STPM) gun with a very nice pattern count of up to 5.000.000 saved pattern. The radar is also something new (a least for me) and is invented with a very high lock count (to collect the pattern quickly) while still getting a good field view. It basically does sweep the radar once than it locks to the nearest target until the other target radar circles reached the bot. Something like a radar wave i guess. This means that i don't scan for targets in the opposite corner for a very long time because they are no thread to me at this position. The implementation is very rude right now because it turned out that this kind of radar has many special rules to handle and i'm to stupid to see how it goes easier. This will be fixed in future releases. The gun is a first draft of my new STPM gun which has its root from the SingleTick bot of Simonton. in its first release the gun works the same like SingleTicks gun but with a very important change. The original gun was not capable to hold much pattern because of very rough memory usage. My data structure can hold a very large amount of visited patterns and as there it has a good base to build on in future releases. The patterns are hold by a map of string hash codes and the pattern ticks. Most pattern are visited only once and it saves huge memory chunks if the data holder is tweaked to this. The movement is the movement of Wallaby and i will use it for the first versions as it is. Many things are open and have to be fixed and changed:

• right now the bot is mega class and it should be mini in the end
• the gun is not ready for wall avoid or collisions or other fancy pattern stuff
• the gun needs some research to find the best values to save the current state (shot at, distance to wall, velocity change and so on).
• the radar has to be smaller and the target search has to be faster (it sweeps right now the hole round even if the target is one tick in the other direction)
• the movement is nice for crowded situations but way to weak to hold the 1v1 challenge and therefore it has to be changed

Credit: Simonton deserves credit for the invention of the SingleTick pattern concept and for the tutorial bot SingleTick
... lets go hunting