From: rars-d-request@lysator.liu.se Subject: rars-d Digest V97 #3 X-Loop: rars-d@lysator.liu.se X-Mailing-List: archive/volume97/3 Precedence: list MIME-Version: 1.0 Content-Type: multipart/digest; boundary="----------------------------" To: rars-d@lysator.liu.se Reply-To: rars@lysator.liu.se (Rars mailing-list) ------------------------------ Content-Type: text/plain rars-d Digest Volume 97 : Issue 3 Today's Topics: Re: Neural Net / GA Drivers Re: Neural Net / GA Drivers Re: Neural Net / GA Drivers Re: Neural Net / GA Drivers Re: Races Results for 10-26 (rules for next season) Re: Races Results for 10-26 (rules for next season) Re: Neural Net / GA Drivers Race Results for 11-9 ------------------------------ Date: Fri, 07 Nov 1997 16:18:12 GMT From: Robert Wilderspin To: "feverish.rars" Subject: Re: Neural Net / GA Drivers Message-ID: <34633e31.57244714@127.0.0.1> On Wed, 05 Nov 1997 13:15:56 -0500, Ralph Scott wrote: >Robert Wilderspin wrote: >> >> A properly chosen neural network is supposed to be able to approximate > >Multilayer neural network? Yeah, with one, or possibly two hidden layers, using back propagation. For a problem as complex as this it's pointless to try a network any simpler than that. >More like it only knew the data for the 3 segs immediately following >where >it was. My original cars were all like this. I should break them out >someday. Horrible trig was used. I've never tried accessing the whole track at once. Three segments should be enough for a decent car though, shouldn't it? How much better can a car get when it knows the entire track? rob. -- Anything that beats my car is obviously cheating. I'll get even one day... -------------------------= (send replies to rob@) ------------------------------ Date: Fri, 07 Nov 1997 16:21:52 GMT From: Robert Wilderspin To: "feverish.rars" Subject: Re: Neural Net / GA Drivers Message-ID: <34633f6d.57561064@127.0.0.1> On Thu, 6 Nov 1997 21:16:34 +0100, kim@ida.his.se (Kim Laurio) wrote: >Just a thought: It could be tough to say "Hey, I'm going to try some >different algorithms on this network and see what happens". So, a >hypothesis that you might want to start with for your thesis work could >be: "Considering the generality of the GA/ES based training, they could >be more appropriate for ANNs where the topology is not fixed", or "I >suspect that there are cases where a general GA/ES performs better than >the tailor-made learning algorithm for that architecture". From that you >could work on to define a set of experiments specifically designed to >test (falsify) your hypothesis. Before you'd even notice it, you'd have >a report in your hands... :-) If only! Unforuntately, the general subject matter of the project has already been chosen, and I have to stick to it. However, for RARS, it looks like GAs are a very good training method, so I'll be doing something with them. They eliminate the need to choose good training data, which is causing me problems with the ANN approach at the moment. >on emergent behaviour in multiple agent systems: >> Why not give it a go yourself? :-) >I just might. If I can find the time for it... :-( That's always the problem... rob. -- Anything that beats my car is obviously cheating. I'll get even one day... -------------------------= (send replies to rob@) ------------------------------ Date: Sun, 09 Nov 1997 10:05:37 -0500 From: Ralph Scott To: Rars mailing-list Subject: Re: Neural Net / GA Drivers Message-ID: <3465D141.2A48@netusa.net> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Robert Wilderspin wrote: > >More like it only knew the data for the 3 segs immediately following > >where > >it was. My original cars were all like this. I should break them out > >someday. Horrible trig was used. > > I've never tried accessing the whole track at once. Three segments > should be enough for a decent car though, shouldn't it? How much > better can a car get when it knows the entire track? Well, the problem I was running into at the end was this. What do you assume about the segment after the ones you know about? You have to assume that the 4th segment following is a wicked hairpin. This may never be true, but for the sake of surviving the track you have to make that assumption. So your car may prematurely be braking based on that assumption. I think I even have a track (of laguna seca) that has this situation. Many cars don't make this assumption and just leave the track. Of course, you could just have a flag for whether a certain track needs that assumption to be made. But then you have started down the path of keeping info on the entire track. ---ralph ------------------------------ Date: Sun, 9 Nov 1997 18:48:14 +0100 (MET) From: Maido Remm To: Rars mailing-list Subject: Re: Neural Net / GA Drivers Message-ID: Content-Type: TEXT/PLAIN; charset=US-ASCII On Sun, 9 Nov 1997, Ralph Scott wrote: > Robert Wilderspin wrote: > > > I've never tried accessing the whole track at once. Three segments > > should be enough for a decent car though, shouldn't it? How much > > better can a car get when it knows the entire track? > Ralph added: > Well, the problem I was running into at the end was this. What do > you assume about the segment after the ones you know about? > You have to assume that the 4th segment following is a wicked hairpin. > This may never be true, but for the sake of surviving the track you > have to make that assumption. So your car may prematurely be braking > based on that assumption. I think I even have a track (of laguna seca) > that has this situation. Many cars don't make this assumption and just > leave the track. > > Of course, you could just have a flag for whether a certain track needs > that assumption to be made. But then you have started down the path > of keeping info on the entire track. I chose absolute coordinate system for calculating angles and distances to certain points on following AND PREVIOUS segments. It is probably possible without knowing all track but at least you have to know trackin[seg].beg_x, trackin[seg].beg_y trackin[seg].beg_ang trackin[seg].end_ang for following segments and save the length and radius of previous segments Can you train your car without knowing that data? Maido PS I think it would be nice to have absolute_xy <-> local_seg_tolft_toend converter function available for every car, this would simplify the car making task for many people who are terrified by converting this data. ------------------------------ Date: Sun, 9 Nov 1997 20:01:05 +0100 (MET) From: Maido Remm To: rars@lysator.liu.se Subject: Re: Races Results for 10-26 (rules for next season) Message-ID: Content-Type: TEXT/PLAIN; charset=US-ASCII > > RACING SCHEDULE (all races track surface 1) > November 9th, Donington (Long) 300 laps > November 23rd, Elkhart Lake (Road America) 300 laps > > TENTATIVE RACING SCHEDULE (all races track surface 1, v.64) > Tentative Alternates: > > In this order... Mosport, Mont Tremblant, Watkins Glen, Zandvoort, > Pocono, Jerez, Albert Park. Jerez? and Albert Park? > -------------------------------------------------------------------- > Comments from Ralph > > The end of the F1 season. Bulle wins in convincing fashion. Followed > by Apex and JAS. Very good job. I was going to try to make an ascii > art rendition of the podium, but I am not very good at that. > > I think there are 4 months until the F1 season starts again. Until then > we will run a series of real world tracks. There are going to several > changes to the rules which I will list here. There will only be one > track per session. 300 miles. The races will be held every two weeks > on the following tracks (found on Maidos www site.) Donington Long, > Elkhart Lake (Road America), Mosport, Mont Tremblant, Watkins Glen, > Zandvoort, Pocono, Jerez, Albert Park. I hope that Jerez and Albert > Park will be finished by that time. In addition I will run the > following drivers along as well Blender, Herbie, Pietro, Ramdu, Sarah, > Tuna, Wappuca3. These will be the cars from the www site as well. I > will run my car as mentioned before, counting for points. IMPORTANT, I > am not automatically carrying over the cars from one season to the next. > You must either resubmit your car or let me know you wish to continue > with your old car. > > Is this your final announcement? I would add this to my web page then. I would like to keep my current Apex8. Nevertheless, I would like to submit my faster but fragile Apex92 also. Can I make two entries or is this forbidden? Current rules do not point it out clearly. I think teams of two will add some thrill (and traffic) to series. Maido ------------------------------ Date: Sun, 09 Nov 1997 14:49:23 -0500 From: Ralph Scott To: Rars mailing-list Subject: Re: Races Results for 10-26 (rules for next season) Message-ID: <346613C3.5C32@netusa.net> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Maido Remm wrote: > > I think there are 4 months until the F1 season starts again. Until then > > we will run a series of real world tracks. There are going to several > > changes to the rules which I will list here. There will only be one > > track per session. 300 miles. The races will be held every two weeks > > on the following tracks (found on Maidos www site.) Donington Long, > > Elkhart Lake (Road America), Mosport, Mont Tremblant, Watkins Glen, > > Zandvoort, Pocono, Jerez, Albert Park. I hope that Jerez and Albert > > Park will be finished by that time. In addition I will run the > > following drivers along as well Blender, Herbie, Pietro, Ramdu, Sarah, > > Tuna, Wappuca3. These will be the cars from the www site as well. I > > will run my car as mentioned before, counting for points. IMPORTANT, I > > am not automatically carrying over the cars from one season to the next. > > You must either resubmit your car or let me know you wish to continue > > with your old car. > > > > > > Is this your final announcement? I would add this to my web page then. It WAS going to be my final announcement. I see now that it may not have been the best choice of alternative drivers. I will use your apex8, and I will accept Apex92. How to use for points?? I don't know. It might be that using two will encourage people to submit their latest but not fully tested cars... I will post to the list and find out. ---ralph ------------------------------ Date: Mon, 10 Nov 1997 11:18:49 +0100 From: kim@ida.his.se (Kim Laurio) To: rars@lysator.liu.se Subject: Re: Neural Net / GA Drivers Message-Id: <199711101018.LAA28715@balder.ida.his.se> > >Robert Wilderspin wrote: > >> > >> A properly chosen neural network is supposed to be able to approximate > > > >Multilayer neural network? > > Yeah, with one, or possibly two hidden layers, using back propagation. > For a problem as complex as this it's pointless to try a network any > simpler than that. Whether you really need multiple layers depends on the nature of the problem. If the set of input vectors is linearly independent, you don't need hidden units. Do you know for a fact that the problem (with the preliminary set of selected input features) is not linearly separable? In other words: Perhaps there is a subset of all the available input features that could be utilized to create a single layer NN that is sufficient for driving a car around the track? Kim ------------------------------ Date: Mon, 10 Nov 1997 18:17:14 -0500 From: Ralph Scott To: rars list Subject: Race Results for 11-9 Message-ID: <346795FA.3FB5@netusa.net> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit RARS Bi-weekly Race Report & News About Upcoming Races November 9, 1997 RARS is the Robot Auto Racing Simulation, a competition for programmers and an on-going challenge for practicioners of Artificial Intelligence and real-time adaptive optimal control. It consists of a simulation of the physics of cars racing on a track, a graphic display of the race, and a separate control program (robot "driver") for each car. All RARS software and activities are free and open to the public. It runs on DOS, Windows, OS/2, UNIX, Linux, and several other platforms. For more information e-mail to mindwarp@cftnet.com or rscott@netusa.net. Or visit the Rars website at http://www.ebc.ee/~mremm/rars/rars.htm Contents Submissions Todays Races Preliminary lap times Race Report Authors Point Standings Upcoming Races Comments -------------------------------------------------------------------- Submissions Please! Any people who want to submit for the first time please give yourselves more than a day or two (I made that mistake when I started). Submissions are to be SOURCE CODE only. -------------------------------------------------------------------- Here is a repeat of the announcement about todays races: November 9th, (Unlimited Division Races) No. of races 1 Tracks: Donington (long) Miles: not less than 300 Software version: 0.64 Surface Type: 1 no practice, random starting order. This is a modified .64 where all cars must cross the finish line once before the race is over. Pitting code has been added. -------------------------------------------------------------------- Race Notes: 10 cars for 120 laps. The track was doningtn.trk. The drivers were: Ralph2, Jas, Rachel01, Apex8, Blender, Burns2, Herbie, Sarah, Tuna, and Ramdu. The initial RVG seed was 11146. results of race 1: starting positions: Herbie, Burns2, Tuna, Blender, Ralph2, Jas, Ramdu, Apex8, Rachel01, and Sarah. 1 Ralph2 avg spd 72.01 120 laps 13761 damage 63 fuel 10 pts accum 2 Apex8 avg spd 71.72 120 laps 11378 damage 39 fuel 6 pts accum 3 Herbie avg spd 70.74 118 laps 12000 damage 85 fuel 4 pts accum 4 Jas avg spd 69.43 116 laps 13413 damage 24 fuel 3 pts accum 5 Blender avg spd 65.58 110 laps 13542 damage 127 fuel 2 pts accum 6 Burns2 avg spd 65.15 109 laps 12000 damage 148 fuel 1 pts accum 7 Sarah avg spd 38.09 64 laps 20862 damage 0 fuel 0 pts accum 8 Tuna avg spd 31 9 laps 30293 damage 149 fuel 0 pts accum 9 Rachel01 avg spd 0 0 laps 30187 damage 147 fuel 0 pts accum 10 Ramdu avg spd 0 0 laps 30024 damage 148 fuel 0 pts accum ---------------------------------------------------------------------- Points for Day Driver Race Last 5 new->old Total ---------------------------------------------------------------------- Ralph 10 10.0 0.0 0.0 0.0 0.0 10.0 Jas 3 3.0 0.0 0.0 0.0 0.0 3.0 Sarah 0.0 0.0 0.0 0.0 0.0 0.0 Apex 6 6.0 0.0 0.0 0.0 0.0 6.0 Blender 2 2.0 0.0 0.0 0.0 0.0 2.0 Burns2 1 1.0 0.0 0.0 0.0 0.0 1.0 Herbie 4 4.0 0.0 0.0 0.0 0.0 4.0 Rachel01 0.0 0.0 0.0 0.0 0.0 0.0 Ramdu 0.0 0.0 0.0 0.0 0.0 0.0 Tuna 0.0 0.0 0.0 0.0 0.0 0.0 ---------------------------------------------------------------------- -------------------------------------------------------------------- This is the point standings. Ralph Scott 10 pts (USA) Ralph2 Maido Remm 6 pts (Estonia) Apex Mario Aleksic 4 pts (USA) Herbie Jurgen Sang 3 pts (Germany) Jas Carl Burke 2 pts (USA) Blender Safai Ma 1 pts (Canada) Burns Daniel Ash 0 pts (USA) Sarah -------------------------------------------------------------------- This is the RARS address book. Juergen Sang jusa@darkstar.bb.bawue.de Daniel Ash deash@primenet.com | Maido Remm mremm@ebc.ee Upcoming RARS Race Meets RACING SCHEDULE (all races track surface 1) November 23rd, Elkhart Lake (Road America) 300 miles December 7th, Mosport 300 miles TENTATIVE RACING SCHEDULE (all races track surface 1, v.64) Tentative Alternates: In this order... Mont Tremblant, Watkins Glen, Zandvoort, Pocono, Jerez, Albert Park. Jerez? and Albert Park? -------------------------------------------------------------------- Comments from Ralph Long race. An updated sarah was introduced but it didn't seem to complete the track like it should. Pietro will not be used, rachel and Ramdu. Don't seem to be good enough for general use. I will try to find a Ramdu that is public and sturdy enough. Herbie turned out to be surprisingly competitive. I will use the wappucar (not wappu3) from the website as or next race. -------------------------------- End of rars-d Digest V97 Issue #3 *********************************