From: rars-d-request@lysator.liu.se Subject: rars-d Digest V97 #2 X-Loop: rars-d@lysator.liu.se X-Mailing-List: archive/volume97/2 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 2 Today's Topics: RE: Car colours in RARS. Re: Neural Net / GA Drivers Re: Neural Net / GA Drivers Re: Neural Net / GA Drivers RE: Car colours in RARS. Re: Neural Net / GA Drivers Re: Car colours in RARS. Re: Neural Net / GA Drivers ------------------------------ Date: Wed, 5 Nov 1997 14:31:55 +0100 From: Torben Thellefsen To: "'Rars mailing-list'" Subject: RE: Car colours in RARS. Message-ID: <01BCE9F7.92CF14D0@PC6> Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 8bit [Daniel Brooks wrote:] [monochrome RARS] >:)And to answer your question: No, never heard of one. Not even grayscale. >I might be making one. I just need to show one car (mine). There is also a >problem with the source I have. It freezes my 386, even when I run it in >no display mode. The colour part is easy: Just give your car {oBLACK,oBLACK} in carz.cpp and the others {oWHITE,oWHITE}, and yours will stand out :-) Tell more about the freeze: Is it when you run your own car or others? >PS: I just had a brainstorm. How bout adding a third mailing list that was >dedicated to mailing the source for RARS? That way every time we geet a >new version, we can put it on the mailer to distribute. it could probably >also serve it up on request. Third? What's the other? Automatic e-mails of the source code can be rather large - the current code weighs in at > 400K uncompressed (including drivers). And you can easily find the current source code on ftp.ijs.com. Patches might be an option, but I think we should keep it to the ftp site and mail announcements here (well, we could also mail the patches themselves here as long as they're also archived on the ftp site). BTW: Who maintains our ftp site? And no, I'm not asking to take over ;-) But the RARS files in /incoming do seem awfully old :-/ Regards, Torben ------------------------------ Date: Wed, 05 Nov 1997 15:29:00 GMT From: Robert Wilderspin To: "feverish.rars" Subject: Re: Neural Net / GA Drivers Message-ID: <34608cc3.3352457@127.0.0.1> On Tue, 4 Nov 1997 17:06:51 +0100, kim@ida.his.se (Kim Laurio) wrote: >Well, I'm not sure that I'm following you here but, isn't ANN just the >architecture (i.e. no learning method specified)? Well, yes, I haven't specified a learning method yet. I was looking for other people's experience with *any* learning methods first. I'll probably start with error back-propagation, as it's fairly simple to understand and implement. Do you suggest any others? I would say that an ANN is both the architecture and the learning method. You can't specify one without the other. If I chose to use counter-propagation, the architecture would have to change also. > I thought it was a >common approach to use a GA to quickly get a decent solution and then use >some sort of gradient based search (which is much faster in pinpointing >the local maxima). That's what my supervisor said. :-) The ANN would act as the gradient based search, using the generalised delta rule, or whatever is appropriate for that network/method. >For an alternative to GA (since you talked about comparing different >learning methods instead of architectures), you could take a look at >Evolution Strategies (ES). There was a short article on a variant of >those in the April issue of Dr. Dobbs (I think?). It was about a method >called "Differential Evolution". Thanks - references like that are what I need! >Assuming that it is sufficient for your thesis to get a car to survive >on most tracks (not necessarily being the fastest), one idea could be >to do some pre-processing for the input to the ANN so that you give >it multiple data of the same type. To understand what I'm thinking of >you could imagine the car having a few range sensors that return the >distance or/and relative speed to obstacles around it. That might be an "advanced" topic. I'd rather avoid heavy abstraction if I can, to start with. >Of course, the interesting case would be the driver that manages to >navigate through traffic. It could be interesting to see the emerging >global behaviour of many such drivers (if successful). This could have >applications in traffic simulations when projecting new highways etc. Why not give it a go yourself? :-) rob. -- Anything that beats my car is obviously cheating. I'll get even one day... -------------------------= (send replies to rob@) ------------------------------ Date: Wed, 05 Nov 1997 15:11:51 GMT From: Robert Wilderspin To: "feverish.rars" Subject: Re: Neural Net / GA Drivers Message-ID: <3460868f.1763978@127.0.0.1> On Tue, 4 Nov 1997 09:40:42 +0100 (MET), Maido Remm wrote: >Initial values are computed from track data. start-x, start-y, and radius >are found by random experimentation and saved if their speed up the car. >Those three define one segment of driving path. The segment ends exactly >where next segment begins( end-x = next_start-x and so on). This is all I >need to know for saving a quick lap. Length and other data are calculated >from these (I am using absolute coordinate system). Sounds good. I don't have the maths skill to do things like that though. :-) >Actually, could you explain what part of driver would you like to cover >by NN. There are many possibilities. My GA car has defined alghoritm for >driving and it uses GA for finding fastest path coordinates on given >track Ideally, the whole car would be modelled by one or two neural nets. That is, no defined algorithms at all, because the exercise is to try and train the nets in such a way that they can model the algorithms themselves. That's a pretty tough goal though. > Scott Nelson and Matt Bucken used NN for finding steering >angle. They had two NNs - one for steering angle and one for velocity. They also tried a case-based approach. >Hope you have more success than they did :-) So do I, although mine isn't a group project, so I'll have to put a lot of effort in. And I have to write a 15,000 word report on it as well, God help me. It's going to be a fun year. :-) rob. -- Anything that beats my car is obviously cheating. I'll get even one day... -------------------------= (send replies to rob@) ------------------------------ Date: Wed, 05 Nov 1997 14:44:33 GMT From: Robert Wilderspin To: "feverish.rars" Subject: Re: Neural Net / GA Drivers Message-ID: <34608202.598498@127.0.0.1> On Mon, 03 Nov 1997 21:16:28 -0500, Ralph Scott wrote: >Robert Wilderspin wrote: >> >> I thought that GAs were good at avoiding local minima or maxima, >> because they have an element of randomness in them? > >I'm not sure if randomness has anything to do with it. Mutation and crossover are both "random" events, in a sense. With any particular crossover there may be a massive change in state for the driver, which would be enough to get it out of a local minima. How do you select which bits to mutate, if not randomly? >I was thinking that perhaps traveling without traffic is best left to >pure computation. Passing, choosing between alternative paths and >stuff would be better suited to ai. Of course, for purposes of your >class, your idea might be easier. A properly chosen neural network is supposed to be able to approximate any mathematical function, I believe. Is that right? If it is, then it should be possible to get a reasonable non-traffic driver. The problem with running in traffic is in getting a good set of training data. GAs may be the only practical solution for this, as that trains them on-the-fly, traffic included. Time to think some more... >> This is ideally what I'm aiming for. A car that only takes in >> information about the track and it's position on that track to produce >> the output vector. We'll see though. That might be too hard. :-) > >This was an idea someone had, have a tourny for cars that do not know >the track (cannot use track data). Do you mean track-specific data, such as "Cornering speed for Stef2 is ..."? Such a general-purpose car is what I dream of, but I'm sure it would be suboptimal in most cases. Done properly, it could adjust things as it was racing, speed up around certain corners, etc. In the real world, of course, drivers *do* set up their cars for each track, based on prior knowledge of the track, so it's not such a bad thing to optimise for one track at a time. >And the real reason I even responded is because there was that one >infamous >web page of someone who did a project just like yours. Offhand I can't >think of it. But altavista should have it if it still exists. I think I know the one you mean. By three lads at Taylor University, Scott Nelson and some others? I contacted them a long time ago, and will do again soon, but they didn't seem to have much success. Still, thanks for reminding me. rob, -- Anything that beats my car is obviously cheating. I'll get even one day... -------------------------= (send replies to rob@) ------------------------------ Date: Wed, 5 Nov 1997 11:45:40 -0500 (EST) From: Daniel Brooks To: Rars mailing-list Subject: RE: Car colours in RARS. Message-ID: Content-Type: TEXT/PLAIN; charset=US-ASCII On Wed, 5 Nov 1997, Torben Thellefsen wrote: :)[Daniel Brooks wrote:] :) :)[monochrome RARS] :)>:)And to answer your question: No, never heard of one. Not even grayscale. :) :)>I might be making one. I just need to show one car (mine). There is also a :)>problem with the source I have. It freezes my 386, even when I run it in :)>no display mode. :) :)The colour part is easy: Just give your car {oBLACK,oBLACK} in carz.cpp and the others {oWHITE,oWHITE}, and yours will stand out :-) :)Tell more about the freeze: Is it when you run your own car or others? That part is easy. I would also have to find where it changes screen modes and change that. Also would have to change the track colors. I would probably use a fill pateren for the grass. I never really tried to figure it out because I was able to use the computer at school. I believe it was probably in my car though. We later found a bug in it that made it crash the other computer as well. It was probably the same thing. :) :)>PS: I just had a brainstorm. How bout adding a third mailing list that was :)>dedicated to mailing the source for RARS? That way every time we geet a :)>new version, we can put it on the mailer to distribute. it could probably :)>also serve it up on request. :) :)Third? What's the other? There is also a digest list, right? :)Automatic e-mails of the source code can be rather large - the current code weighs in at > 400K uncompressed (including drivers). And you can easily find the current source code on ftp.ijs.com. :)Patches might be an option, but I think we should keep it to the ftp site and mail announcements here (well, we could also mail the patches themselves here as long as they're also archived on the ftp site). yea, so. I wouldn't mind a 400k dl, especially since I don't have to mess with the ftp site. Can you really patch source code? TTYL, Daniel Brooks ________________________________________________________________________ __ Daniel Brooks - d-brooks@usa.net, dbrooks@geocities.com / /\ http://www.geocities.com/SiliconValley/Vista/2083 / / \ / / /\ \ HANDLE WITH EXTREME CARE: This Product Contains Minute / / /\ \ \ Electrically Charged Particles Moving at Velocities in / / / \ \ \ Excess of Five Hundred Million Miles Per Hour. / /_/____\ \ \ /__________\ \ \ "We're very sorry, Mister Schrodinger, \_____________\/ but the cat refuses to go in the box." ------------------------------ Date: Wed, 05 Nov 1997 13:15:56 -0500 From: Ralph Scott To: Rars mailing-list Subject: Re: Neural Net / GA Drivers Message-ID: <3460B7DC.2212@netusa.net> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Robert Wilderspin wrote: > > On Mon, 03 Nov 1997 21:16:28 -0500, Ralph Scott > wrote: > > >Robert Wilderspin wrote: > >> > >> I thought that GAs were good at avoiding local minima or maxima, > >> because they have an element of randomness in them? > > > >I'm not sure if randomness has anything to do with it. > > Mutation and crossover are both "random" events, in a sense. With any > particular crossover there may be a massive change in state for the > driver, which would be enough to get it out of a local minima. How do > you select which bits to mutate, if not randomly? I meant more that I didn't think that GAs were good at avoiding minima deu SOLELY to the fact that they have AN ELEMENT of randomness. I mean they could be non random somehow and also avoid minima, and not all random GAs would necessarily avoid the minima. > > >I was thinking that perhaps traveling without traffic is best left to > >pure computation. Passing, choosing between alternative paths and > >stuff would be better suited to ai. Of course, for purposes of your > >class, your idea might be easier. > > A properly chosen neural network is supposed to be able to approximate Multilayer neural network? > >This was an idea someone had, have a tourny for cars that do not know > >the track (cannot use track data). > > Do you mean track-specific data, such as "Cornering speed for Stef2 is > ..."? Such a general-purpose car is what I dream of, but I'm sure it > would be suboptimal in most cases. Done properly, it could adjust 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 think I know the one you mean. By three lads at Taylor University, > Scott Nelson and some others? I contacted them a long time ago, and Yes. That was them. ---ralph ------------------------------ Date: Wed, 05 Nov 1997 13:38:29 -0500 From: Jim Sokoloff To: Rars mailing-list Subject: Re: Car colours in RARS. Message-ID: <3460BD25.76C1@tiac.net> Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Daniel Brooks wrote: > yea, so. I wouldn't mind a 400k dl, especially since I don't have to mess > with the ftp site. The issue is estimating the efficiency of "pushing" code to everyone (by e-mail) or letting the people who want a particular piece to "pull" it down to them using ftp (*FILE TRANSFER* protocol :-) ) ftp isn't rocket science; about the only (minor) advantage I can see to e-mail over ftp is that not everyone has a good internet connection to the ftp site, and almost everyone has a good connection to their mail destination. (However, the advantages of ftp are many, IMO, and if the ftp "locality" is a real problem, it's better to solve that by mirroring than by abusing e-mail to accomplish file transfers... :-) ) ---Jim ------------------------------ Date: Thu, 6 Nov 1997 21:16:34 +0100 From: kim@ida.his.se (Kim Laurio) To: rars@lysator.liu.se Subject: Re: Neural Net / GA Drivers Message-Id: <199711062016.VAA22168@balder.ida.his.se> rob wrote: > I would say that an ANN is both the architecture and the learning > method. You can't specify one without the other. If I chose to use > counter-propagation, the architecture would have to change also. Ok, I see your point. 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... :-) 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... :-( Kim -------------------------------- End of rars-d Digest V97 Issue #2 *********************************