Talk:RoboResearch

If you update from SVN, be ready to change your .cfg files. The example "run.cfg" is updated - it's still easy. Perhaps easier than they were! --Simonton 04:25, 16 September 2008 (UTC)

I finally put this up on GitHub, I know a lot of us have moved over to RoboRunner/Jogger, but it doesn't quite feel as well polished as RoboResearch. Also this way we can all put our hacks in one place.

RoboResearch

On that note I have made a new build. It has a new and shiny Challenge Creator function now, right in the program. I still don't have all the new features RoboRunner added to the rrc format yet. In fact a spec should probably be decided one at some point. But I digress.

It still has some bugs I have to work out. I know a lot were reported here, but I somewhat lost track of which are fixed and which are not.

Any chance you feel like elaborating on what you like about RoboResearch over RoboRunner? It would take a heck of a lot to convince me to waste 25% of my CPU time on JVM restarts again. From the command line, I honestly feel like RoboRunner is superior in every way. I haven't used the RoboResearch GUI or RoboJogger much though.

On that I agree. RoboRunner is definitely better on the command line. However I tend to live in the GUI world, and I feel it's GUI is quite a bit ahead of RoboJogger.

Though I now consider I could have just made my own front end for RoboRunner. Since that would probably be much easier then trying to hook it into RoboResearch.

Eh... I'm sure plenty of the RoboResearch UI code could be applicable to a RoboRunner frontend.

Just to chime in on the UI comparison. I recently came back to Robocode, and having been using the RoboResearch GUI in the past, I decided to try RoboRunner w/ RoboJogger. While I like RoboRunner's performance, I prefer the RoboResearch GUI over RoboJogger... but I think I'll end up going with RoboRunner CLI now, as I don't normally live in the GUI world anyway.

Sorta off-topic, but some food for thought...

I took what I think is a somewhat novel approach to the GUI for the automation API in BerryBots. As background, you're writing a program that is run by BerryBots, like a bot - as opposed to Robocode, where you're writing a stand-alone program that embeds the Robocode engine. A big reason for this is so you can write automation code in the same language as the bots. But it also lets me do some of the heavy-lifting needed by most automation code, like multi-threading and user inputs.

So the Game Runner API provides the GUI/input functionality. You get a RunnerForm object when your Game Runner executes. You set what inputs you want, you decide when to show the form to the user, and after that you can fetch the values. The engine dynamically creates the appropriate GUI, shows it to the user, blocks until they hit OK/Cancel, and then the Game Runner resumes execution. You can see an example in simpletourney, which lets you run a tournament. I could also present a fully text-driven form, like if/when I write an ncurses UI.

I'd like to also provide the flip-side, so you could easily write a stand-alone program that uses BerryBots. That's actually not far off from the binary in use for the web UI. The script hit by the web UI is a stand-alone program that runs a BerryBots match from the command line and processes the results.

Lastly, it might also be cool to write a Game Runner-ish API for Robocode, maybe on top of RoboRunner. Something like running a Twin Duel tournament couldn't really be specified as a challenge file, but it would benefit from the multi-threading and some of the score details stuff that RoboRunner already has. Note that simpletourney offers the same tourney functionality as Twin Duel/Tourney Runner, but it's less code, everything's configurable at run-time via a GUI, and matches run multi-threaded.

RoboResearch is not starting JVMs in debug mode. It passes "-Ddebug" as parameter, but it should be "-Ddebug=true".

So that's just an argument to the JVM, right? Custom JVM args is near the top of my list for RoboRunner, so I'm working on it now. But there are some things like "-nodisplay" that are arguments to Robocode itself, which wouldn't work because I don't launch Robocode (just use the control API from my own program). So I'd have to add arguments myself if I wanted to let you configure it to actually display battles, for instance.

Nevermind, did some testing. =)

The bug is in RoboResearch, not RoboRunner. Currently I'm using a custom BattleRunner class which passes the correct argument. It's makes a 2% APS difference in some challenges when using all cores.

I'll try RoboRunner someday.

Yeah, I know, it just reminded me of the related issue in RoboRunner. It's funny with RoboResearch, I feel like everyone has their own set of hacks in their local instance. =)

Wow, 2% difference in debug mode? That's crazy.

I'm strongly considering writing up a new Robocode battle runner to replace my RoboResearch usage. I think I could even have something workable over the weekend. Here's my short list of design goals:

• Easy to setup / configure!
• Use Robocode control API instead of running a fresh Robocode engine / JVM from the command line for each battle.
• Support existing .rrc files.
• Support any battle configuration (battle size, Melee, Twin Duel, etc).

Anyone have any thoughts / requests? I think I would go command line only to start, but maybe if I really get something good I'll consider a GUI.

Now that I can crank out 1,000 battles per hour, the thought of also restarting the JVM 1,000 times per hour kind of makes me cringe. =)

You can take for base Distributed_Robocode

I really should take a look at what you have there, I've never tried it. Though I think I'm looking for something a bit simpler and lightweight. Does the client do multiple threads with multiple Robocode engines? Or do you just have to configure one client per thread, like with RoboRumble clients?

I just whipped up the core battle runner, which takes paths to multiple Robocode install directories, then runs a collection of battles across the threads using the control API. Though that's pretty much the easy part... =)

It uses one thread per server.

Yes, but in my case, i have more CPU power at my work, so support of remote servers was critical for me:) Latter today, i will publish soures on github.

Done

Yeah, support for remote clients is a killer feature in that situation!

Great feature list. Support of current .rrc files is a good idea. You'd need some sort of config file to parse anyway, and it's a nice, easy to read config format anyway.

Remote servers would be fun for me too. I've got my main home desktop, and an identical machine running as our living room media center.. so both of them could run battles for me. *evil grin*

I said it over in my talk page, Voidious, but I'd love to help contribute to this with code and testing. As a teacher my free time will contract somewhat during the school year.

My recent experience with RR has me interested in helping to make something a bit easier to get off the ground.

Hmm. It seems that Robocode itself is not thread safe for running multiple instances? I have multiple RobocodeEngine's going in separate threads, each with its own directory, and I'm still hitting some weird concurrency issues. Bummer! I wonder if I can do something with ClassLoaders to deal with it? I've never really played with them before, and not sure if it would kill performance or not.

Oh, yes i also faced with this problem and also think about class loaders but in result came to current architecture

Yep, the engine is not thread safe.

Creating ClassLoaders should be faster than creating separate JVMs, but don't know by how much.

Another approach would be staying with separate JVMs, but not restarting them before each battle. Exchanging messages between JVMs and using the API inside a loop instead, much like how the RoboRumble client works.

Creating ClassLoaders once and reusing them on many battles is also another approach. I remember trying this approach in a custom multi-threaded RoboRumble client a long time ago. Didn't remember exactly what went wrong. Maybe the need of a custom JAR loading mechanism.

Yeah, having each engine in its own process and doing IPC seemed like a huge ordeal that I wanted to avoid, but now I'm thinking it might not be so bad, and it would be easier and safer than fussing with classloaders. I brushed up on some RMI stuff last night, but even that requires some setup to get the RMI registry going and stuff, which I don't really want to force on people. I think just starting the separate processes with Runtime.exec and communicating with them over STDIN and STDOUT may be the simplest solution all around.

Cool, that wasn't so hard. =) I have this pretty much working now, including reading .rrc files, with option additional stuff like multiple bots per line for Melee battles and custom battle field sizes, and simple/fast/human readable results output to .properties files instead of a database. Still a few things left to do for usability, which is a high priority for me, but I think I can have something posted soon.

~29 mins for one season of a 500-bot tested, vs ~40 mins with RoboResearch. Yum. =)

Niice!

I've almost got deBroglie's score on the testbed you made for me back to the level of rev108 after my hitrate normalization fixes. It's been a great boon to have a working batch battle runner!

Looking forward to testing this work of yours out!

Sounds awesome. When there get to be a lot of robots in the install directory RoboResearch gets really slow starting each battle... try running a single season against the entire rumble and you'll see what I'm talking about.

I also see big slowdowns once the database gets too big, say 15k battles, which is a pretty frequent occurrence for me these days. It's pretty obvious when the first 6 battles fire right up at startup vs trickling in with half a second between them.

I'm not sure I'll have time to get it polished and posted tonight, but if not, definitely within the next day or two. And I'll post it on GitHub as public domain. It's not even that much code, if anyone wants to learn it to fork it or add a GUI or whatever.

There will still be a bunch of bells and whistles left on the to-do list as of the first release, but for general APS bot benchmarking it's in great shape. I'm using it for my own battles now and it ran overnight with no issues. I think I have the setup process pretty smoothed out, too (albeit Unix only, using a bash script).

Voidious, i like yours solution and think, that 3 tools for batch battles execution is too much. But i still need in execution of battles on remote machines, so i think it will be cool to combine some way of our solutions. Rigth now i have no inspiration for work on it, but if you make good interface for yours tool, i can add my client-server solution in future. What do you think? We can contact via email or skype to discuss interface.

Well, since my new tool is more of a replacement for RoboResearch, that only leaves us with 2, and I don't think 2 is too many. =) If you're only using one machine, then a fast, powerful tool that's easy to use for multi-threaded benchmarks is great. If you have multiple machines, no matter what you do it's going to require some setup, so having a tool for that is great too. My multi-process BattleRunner is pretty simple, though, so it shouldn't be hard to integrate with that if we want to use that in the client part of your system.

Problem in my solution is that you need run several servers on same machine to utilize cpu completly and i wish to run only one server, which will run several engines. Also even 2 is too much. It means that every feature (ui, hit rate colculation, melee support etc.) must be doubled and i think, that better for community to concentrate all resources on one tool. I will review your tool, when you publish it and contact you, if detect troubles in integration or impruvement ideas.

Is Chase leading the support for RoboResearch development now? Just let me know if you would like extra help. I have plenty of experience in Java and Swing.

I might end up redoing a bit of the GUI when I have the chance. But I don't even have time to really work on my robots at the moment, let alone on RoboResearch. :)

On that note I also want to switch it over to use SQLite instead of what it currently uses.

Doing both of these is not hugely difficult, the main problem will be transitioning the current robocode control and challenge definitions over, etc.

With my latest versions of XanderCat 9.0 and 9.0.1, they seem to perform just fine with 1.7.3.0, but I am now getting really wacky scores in RoboResearch. There are no exceptions happening, but the scores are coming up a good 20% to 30% lower than the scores with 1.7.3.0 (e.g. score against Rapture.Rapture is coming up at about 65% in RoboResearch, but a consistent 90% in 1.7.3.0 test runs and in the RoboRumble). I checked to ensure my robot is not throwing any exceptions. Any thoughts on what might be causing this? It is preventing me from using RoboResearch for testing.

It might be due to me setting it to debug mode (it keeps robots from skipping turns, so if enemies were skipping, they are not in RR).

I don't think so. I might have to spend some time digging into the code more. A lot of the docs and pages discussing RoboResearch seem to be outdated. For example, I keep seeing references to having a copy of Robocode installed in the robocode_install directory, but from what I can see, there doesn't seem to be any robocode_install directory. I assumed it just used robocode.jar from the lib directory. Quite confusing. So I don't even know what version of Robocode RoboResearch is using.

AFAIK you're supposed to install a copy of Robocode into robocode_install. But the original RoboResearch also came with a lib/robocode.jar that it used, which I only just recently realized it uses instead of the one in robocode_install/libs. I'm not sure what the problem is, but I'd install Robocode 1.7.3.x to robocode_install, start it up and recalculate the CPU constant, and copy that robocode.jar into roboresearch/lib. See if you still see a score discrepancy.

RoboResearch docs could definitely use a lot of work, but hopefully with Chase taking charge to start updating it again recently, we'll resolve some of the usability issues and keep it better maintained.

I dug into the code a little to figure it out. I didn't realize that RoboResearch just runs Robocode normally using "-nodisplay". That made it easy to debug, because I could just comment out the line that added "-nodisplay" and watch the battle as it ran. The version of Robocode in the lib directory that it is using is 1.6.1.2. For whatever reason, it appears as though the onRoundEnded() event never gets called, which is messing up some of the round data I keep track of. Ultimately, this lead to one of my multi-mode scenarios getting incorrectly activated much more often than it should have been, resulting in low scores. Now I need to either develop a work-around or try to set up RoboResearch to use 1.7.3.0.

Ah, cool. I wondered if it might be some robot method added in 1.7.x, which I think onRoundEnded is. You should be able to copy robocode.jar from robocode/libs to roboresearch/lib to get whatever Robocode version you want.

I modified my RoboResearch to use robocode_install/libs/robocode.jar, so whichever Robocode version I install there is used. That and the other patches I'm using are available here, and described/linked a bit further up this talk page. Recompile with javac -cp src:hsqldb.jar -d bin src/roboResearch/CLI.java (or GUI.java).

Oh yeah, there may be some other minor incompatibilities using 1.7.x. I fixed some issues with the command line version of RoboResearch in my patch, and Chase may have fixed some with the GUI in his version (if there were any).