In academia, it's more or less mandatory to have a good homepage, with all publications downloadable and a complete CV, at the least. Researchers without a homepage arouse suspicion, for good reason: what good is your research if nobody knows about it?
So yes, it definitely helps to post a lot about yourself online if you want to get an academic job.
I disagree with the conclusions of Mayer-Schönberger, though I agree with the main logic of his argument.
I believe making all public archives "forgetful" would be (A) disastrous to research in history, philology, and linguistics (e.g.) and of course to political accountability, and (B) almost impossible, at least without a technological monoculture (e.g. Microsoft runs all the blogs).
Instead, we have to adapt our culture to the inevatible presence of modern technology. This means that if someone once made racist, paedophiliac, hateful, misogynic, androgynic, stupid, schizofrenic etc. remarks, this should not in any way be held against them when they later in life want to become a politician, teacher, babysitter, policeman etc. We will simply have to assume that people can change and restrict ourselves to looking at their most recent behaviour and opinions.
Changing our culture in such a way might sound impossible (and to some people undesirable), but I think it's far more possible (and desirable) than changing our technology in the way Mayer-Schönberger proposes.
It is also possible that such a cultural change would be a natural consequence of information about everybody becoming available, rather than it making us all into politicians.
"Driving game AI hasn't improved much but frankly that's because there's not a lot to think about with driving games."
If you were right in the second part of your statement, everything I've been doing in the last two and a half years would be completely meaningless. As I don't want to have wasted these years, I prefer to think that you are not right.
There's a lot to do about driving game AI. First of all, learning to drive well on complicated tracks - without cheating - is not at all straightforward. Keeping the same performance when the user is allowed to create his own tracks is even harder - most racing games rely on knowing their prefabricated tracks well, tracks which are made from a set of standard segments in order to be tractable for the AI. When you introduce more than one car on the track it gets even trickier, as you have to deal with overtaking, collision avoidance, forcing collisions, etc.
And these are just the challenges associated with generating good driving. Interesting driving has even more challenges - should you drive nicely or aggressively? How do you make the driving look human-like? How do you adapt your skill level to that of your opponents? Etc...
You can see this blog post (with videos) and this paper for some of the research we are doing into this.
Wow, that is a lot of interesting comments, from someone in the know as well... I am very happy to get people other than computer scientists commenting on my research!
First of all, I know nothing about real car racing. I don't even have a driver's license, that's how bad it is. (To my defence, not every Swede has one...)
This particular paper is very game focused, true. As such, it does not deal with noisy sensors and all that, but we do have another paper on how to model real (miniature R/C) car dynamics, evolve controllers in the model and transfer them back to the real car, bringing it all closer to the real world. It is currently being revised and will be up on the website within a week or so.
Now to your actual comments. We chose no to rely on standard segments and instead on b-splines (or, in the earlier workshop paper, a form of parametrised segments) to allow evolution more room for its creativity. We explicitly wanted evolution to come up with things we had never seen before. But if it come up with features which would not have the intended effect on real drivers - such as differences in the width of the track that a human driver would just not care about - this is a problem. In that case it is a problem with our fitness function, thus probably with the underlying player model. We are aware of the need for more accurate player models based on more complex sensory inputs. And of doing all this with multi-car racing.
One alternative would be to use an existing racing game such as Trackmania which comes with a track editor based on segments but which still allows for a lot of creativity in constructing tracks.
Now that's an interesting point. Hmm... it would require internally modelling the field of attention, and judgement, of the judge. Or just copying what the other cars do. But I agree, it makes the problem harder - and more interesting!
According to my code, the minimum turning radius is 25 pixels. (The length of the car is 20 pixels and width 10 pixels.) But that's very minimum, at the speeds it actually travels the actual turning radius is quite a bit larger. I've never cared to calculate how much, as the simulation is not supposed to quantitatively model anything. However, we have been working recently on automatically modelling a small radio-controlled car, and evolving controllers based on that model that then perform well in the real world. I'll put that paper up on the website within the next few days.
That's interesting. I have to get back to you about the minimum turning radius once I get to my "work" computer (just woke up). I'd like to know about the results!
This is an interesting argument, and yes, we are experimenting with predictive control and internal models as well (experiments running as we speak). But I think that the real power of predictive models only show when you have multiple agents involved. People often underestimate the power of reactive control just because they assume that humans can't tune them well enough. But evolution is not limited in that way, and can sometimes do things with reactive controllers you just wouldn't think possible. (Stefano Nolfi has done some thinking and experimenting on this, see for example his paper "Power and limits of reactive agents".)
In the case of our (admittedly simple) model, we have a limited line of sight, and I think a good reactive controller can perform optimally (however you define that, often optimality is just another buzzword) given the limited sensor data. We did try evolving reactive control on much narrower tracks with good results - see for example our papers on track evolution. What the controller learns is often just to slow down when coming up to a narrow passage.
Thanks for the nice comments! It's all very free and open source, though I haven't gotten around to spelling it out explicitly or attaching any particular license to it. I too would really like to have a more sophisticated simulator to work with, but I have two really important requirements which all the alternatives I looked at so far have run afoul of:
* The simulator should be 100% cross platform (I regulary use Ubuntu, Mac and Windows, and researchers who want to participate in the competition can well be expected to run more exotic systems). This pretty much rules out anything written in c, as the graphical frontend would be platform specific.
* It should be easy to get started with. Really, really easy. Even if I might be willing to spend a week learning the systems, researchers who want to participate in the competition will certainly. We're talking about something like five minutes here, provided you're proficient in programming in general (but not necessarily the specific language chosen).
Someone higher up mentioned TORCS (or was it RARS? I keep forgetting which is which...) and while it seems nice enough a simulation, with quite a user base, it is clearly not platform agnostic and easy to get started with enough.
Has anyone got a good suggestion? Otherwise, there is some serious development work to be done here, at some point...
If anyone actually is interested in reading the papers discussing the experiments we did (many more than you see in the videos!), most of them are available on my website.
Some of them are of course better than others. I can recommend this one, about evolving general and specific driving skills, this one about co-evolution, this one about different learning techniques, and this one about modelling human driving and evolving tracks. There are several new ones, including one on physical cars, which are not on the website yet - mail me if you want a preprint!
All this assuming that anyone actually reads academic papers... sometimes it seems that not even the guy who writes the paper actually reads it. (Not true in my case, of course!)
I don't know much about the techniques underlying Forza 2, but I went over and talked to the guys who worked on Forza 1, and we compared our approaches. At least for the first game, what they are actually using is recorded trajectories on different track segments which are then spliced together at the junctions of segments, so as to create similar-looking behaviours on unseen tracks. The problem here is of course that the new tracks are constrained to being constructed out of the same segments as the driver has already been tested on - there is no generalization. The track designers for Forza simply had to live with this constraint.
We have ourselves gotten player modelling working fine with evolutionary neural networks, which can generalize, but the Forza team didn't consider these techniques reliable and fast enough in time for the release of the original game. Maybe things have changed with Forza 2.
Oh yes. I was stupid enough to move some domains onto GoDaddy's servers, without checking out how their email forwarding works. It turns out they have a blacklist you can't opt out of. Very very annoying, to say the least - it's absolutely essential for any thinking being to be able to see what your spam filter lists as spam and correct it on occasion.
Yes, I will change hosting provider. As soon as I get the time. But I would of course prefer GoDaddy changing to being reasonable, and at least allowing me to opt out of having their blacklist applied to all my incoming mail.
I don't see anything wrong with getting paid for mentioning / exhibiting brands (or even arguments) in your art. This is already commonplace in movies, and has been done on occasion by high-profile musicians and novelists.
Given that most people seem to have little against advertising, as long as the ads are good (but typically has a lot against many very bland commercials that litter our mass media), this seems like a win-win situation to me.
...it might not be very useful in itself, but it can fruitfully combined with almost anything. So a combination of maths and something applied is a very good idea. So lots of mathematics with some computer science is a very good idea - probably better than the other way around, from a advanced jobhunting point of view. The specifics of various algorithms and development environments are easy to pick up later, while the a proper mathematical background had better be there from the start.
If you have time over, try to throw in some other random sciences as well - a bit of physics? Some neuroscience? A short course in geology? Broadening your view is always useful.
That you don't want to work with computers is a problem, as that is what you will end up doing in virtually any science or engineering position. So get over it.
Further, I agree with above poster that if you're thinking about getting a PhD you should not wait too long. Getting used to having a proper salary is a major demotivator. So take maximum a year out of university.
The problem you mention, that you don't know exactly how the system will behave under all possible conditions, is a problem we have with all computer programs, especially those that include learning. On the other hand, this is a problem we have with humans as well. The reason (well, one of them) that we let humans drive is that we have done such extensive testing with humans driving, to see under which conditions they can drive safely and under which conditions the "behave unexpectedly" (icy roads, fog, weird intersection layouts, poorly marked turns etc). In the process, tens of thousands of humans drove themselves to death.
So I wouldn't count out self-driving cars until we have done the same amount of testing with them.
Another way we can try to make self-driving cars safer is to use something like Rodney Brooks' Subsumption Architecture, where the controller is structured into layers. The higher layers are responsible for "high-level" behaviour like navigating and planning fuel consumption, while lower layers do simpler things like avoid driving into walls based on sonar information. If a higher layer falls, the lower layers operate just fine without it. The function of the very lowest layer could then be just to stop the car if all the higher layers fail. Given that the layers are separate circuits, it's very unlikely that all will fail simultaneously.
Much of the design can be automated. I'm myself working on using evolutionary algorithms to design car controllers. So far we do quite simple things and only in simulation, but with a good enough simulation it could then be transferred to a real car. Check out this post, including videos.
Stefano Nolfi's team (responsible for the research in TFA) usually come up with very solid and ingenious research, and the project mentioned is a continuation of a large EU-funded project with interesting results.
But it should be remembered that this research is quite far from any actual applications, in large part because people schooled in classical engineering instinctively distrust anything that has been "evolved". Understandable, as there is often no way to tell why the evolved AI is doing what it is doing, just that it is doing it, and no guarantees that it will continue doing it when you need it the most.
Sometime in the far future, however, I believe artificial evolution might be the standard way of creating control and intelligence for various purposes.
While computing power is of course good have, I don't think the computing power of individual agents is a major factor hindering development in this type of AI. Our own research (which is similar in spirit to Nolfi's, from TFA) points to the importance of appropriate sensor setups, environments and tasks, and that much can be done with simple neural networks. (On the other hand, if you work in simulation, much processing power might be needed for simulating the environment of the robot.) I recently wrote an article about this, and how computer games might provide the appropriate tasks and environments.
I don't agree. We basically only need to reward the robots appropriately (just like the zombies in "day of the dead") for artificial evolution to create the intelligence for us. "Wants" and "needs" are just words we use to label certain cognitive mechanisms, out of similarity with how we perceive our own thinking. If having "wants" and "needs" leads to better fitness (higher rewards) for the robot, evolution will come up with those things.
I just wrote a post describing the general idea behind this approach to artificial intelligence - check it out!
Slashdot Mirror
In academia, it's more or less mandatory to have a good homepage, with all publications downloadable and a complete CV, at the least. Researchers without a homepage arouse suspicion, for good reason: what good is your research if nobody knows about it? So yes, it definitely helps to post a lot about yourself online if you want to get an academic job.
This is pretty theoretical, and arguably quite fundamental: http://www.idsia.ch/~juergen/unilearn.html Disclosure: I work with Juergen Schmidhuber
> As it is, I use Safari on Mac and Firefox on PC and Firefox is by far the more comfortable in use.
So why don't you use Firefox on the Mac as well?
I disagree with the conclusions of Mayer-Schönberger, though I agree with the main logic of his argument.
I believe making all public archives "forgetful" would be (A) disastrous to research in history, philology, and linguistics (e.g.) and of course to political accountability, and (B) almost impossible, at least without a technological monoculture (e.g. Microsoft runs all the blogs).
Instead, we have to adapt our culture to the inevatible presence of modern technology. This means that if someone once made racist, paedophiliac, hateful, misogynic, androgynic, stupid, schizofrenic etc. remarks, this should not in any way be held against them when they later in life want to become a politician, teacher, babysitter, policeman etc. We will simply have to assume that people can change and restrict ourselves to looking at their most recent behaviour and opinions.
Changing our culture in such a way might sound impossible (and to some people undesirable), but I think it's far more possible (and desirable) than changing our technology in the way Mayer-Schönberger proposes.
It is also possible that such a cultural change would be a natural consequence of information about everybody becoming available, rather than it making us all into politicians.
"Driving game AI hasn't improved much but frankly that's because there's not a lot to think about with driving games."
If you were right in the second part of your statement, everything I've been doing in the last two and a half years would be completely meaningless. As I don't want to have wasted these years, I prefer to think that you are not right.
There's a lot to do about driving game AI. First of all, learning to drive well on complicated tracks - without cheating - is not at all straightforward. Keeping the same performance when the user is allowed to create his own tracks is even harder - most racing games rely on knowing their prefabricated tracks well, tracks which are made from a set of standard segments in order to be tractable for the AI. When you introduce more than one car on the track it gets even trickier, as you have to deal with overtaking, collision avoidance, forcing collisions, etc.
And these are just the challenges associated with generating good driving. Interesting driving has even more challenges - should you drive nicely or aggressively? How do you make the driving look human-like? How do you adapt your skill level to that of your opponents? Etc...
You can see this blog post (with videos) and this paper for some of the research we are doing into this.
Wow, that is a lot of interesting comments, from someone in the know as well... I am very happy to get people other than computer scientists commenting on my research!
First of all, I know nothing about real car racing. I don't even have a driver's license, that's how bad it is. (To my defence, not every Swede has one...)
This particular paper is very game focused, true. As such, it does not deal with noisy sensors and all that, but we do have another paper on how to model real (miniature R/C) car dynamics, evolve controllers in the model and transfer them back to the real car, bringing it all closer to the real world. It is currently being revised and will be up on the website within a week or so.
Now to your actual comments. We chose no to rely on standard segments and instead on b-splines (or, in the earlier workshop paper, a form of parametrised segments) to allow evolution more room for its creativity. We explicitly wanted evolution to come up with things we had never seen before. But if it come up with features which would not have the intended effect on real drivers - such as differences in the width of the track that a human driver would just not care about - this is a problem. In that case it is a problem with our fitness function, thus probably with the underlying player model. We are aware of the need for more accurate player models based on more complex sensory inputs. And of doing all this with multi-car racing.
One alternative would be to use an existing racing game such as Trackmania which comes with a track editor based on segments but which still allows for a lot of creativity in constructing tracks.
Now that's an interesting point. Hmm... it would require internally modelling the field of attention, and judgement, of the judge. Or just copying what the other cars do. But I agree, it makes the problem harder - and more interesting!
According to my code, the minimum turning radius is 25 pixels. (The length of the car is 20 pixels and width 10 pixels.) But that's very minimum, at the speeds it actually travels the actual turning radius is quite a bit larger. I've never cared to calculate how much, as the simulation is not supposed to quantitatively model anything. However, we have been working recently on automatically modelling a small radio-controlled car, and evolving controllers based on that model that then perform well in the real world. I'll put that paper up on the website within the next few days.
That's interesting. I have to get back to you about the minimum turning radius once I get to my "work" computer (just woke up). I'd like to know about the results!
This is an interesting argument, and yes, we are experimenting with predictive control and internal models as well (experiments running as we speak). But I think that the real power of predictive models only show when you have multiple agents involved. People often underestimate the power of reactive control just because they assume that humans can't tune them well enough. But evolution is not limited in that way, and can sometimes do things with reactive controllers you just wouldn't think possible. (Stefano Nolfi has done some thinking and experimenting on this, see for example his paper "Power and limits of reactive agents".)
In the case of our (admittedly simple) model, we have a limited line of sight, and I think a good reactive controller can perform optimally (however you define that, often optimality is just another buzzword) given the limited sensor data. We did try evolving reactive control on much narrower tracks with good results - see for example our papers on track evolution. What the controller learns is often just to slow down when coming up to a narrow passage.
Thanks for the nice comments! It's all very free and open source, though I haven't gotten around to spelling it out explicitly or attaching any particular license to it. I too would really like to have a more sophisticated simulator to work with, but I have two really important requirements which all the alternatives I looked at so far have run afoul of:
* The simulator should be 100% cross platform (I regulary use Ubuntu, Mac and Windows, and researchers who want to participate in the competition can well be expected to run more exotic systems). This pretty much rules out anything written in c, as the graphical frontend would be platform specific.
* It should be easy to get started with. Really, really easy. Even if I might be willing to spend a week learning the systems, researchers who want to participate in the competition will certainly. We're talking about something like five minutes here, provided you're proficient in programming in general (but not necessarily the specific language chosen).
Someone higher up mentioned TORCS (or was it RARS? I keep forgetting which is which...) and while it seems nice enough a simulation, with quite a user base, it is clearly not platform agnostic and easy to get started with enough.
Has anyone got a good suggestion? Otherwise, there is some serious development work to be done here, at some point...
Hmm... why? Why would a machine learning systems never be able to learn such moves? And why wouldn't they be able to stick to the rules? Just curious.
If anyone actually is interested in reading the papers discussing the experiments we did (many more than you see in the videos!), most of them are available on my website.
Some of them are of course better than others. I can recommend this one, about evolving general and specific driving skills, this one about co-evolution, this one about different learning techniques, and this one about modelling human driving and evolving tracks. There are several new ones, including one on physical cars, which are not on the website yet - mail me if you want a preprint!
All this assuming that anyone actually reads academic papers... sometimes it seems that not even the guy who writes the paper actually reads it. (Not true in my case, of course!)
...and for once, this important questions is actually answered: these neural networks are running on Linux right as I write this!
I don't know much about the techniques underlying Forza 2, but I went over and talked to the guys who worked on Forza 1, and we compared our approaches. At least for the first game, what they are actually using is recorded trajectories on different track segments which are then spliced together at the junctions of segments, so as to create similar-looking behaviours on unseen tracks. The problem here is of course that the new tracks are constrained to being constructed out of the same segments as the driver has already been tested on - there is no generalization. The track designers for Forza simply had to live with this constraint.
.
We have ourselves gotten player modelling working fine with evolutionary neural networks, which can generalize, but the Forza team didn't consider these techniques reliable and fast enough in time for the release of the original game. Maybe things have changed with Forza 2.
There is some information on the Forza AI on http://research.microsoft.com/mlp/forza/, and our approach to modelling is described in http://julian.togelius.com/Togelius2006Making.pdf
Note that all this is about modelling behaviour, not about creating new behaviour from scratch; there are some papers on this on my website as well.
> your post was marked +5 informative yet has miss information in it
Miss information! Where? I've always wanted to meet her!
Oh yes. I was stupid enough to move some domains onto GoDaddy's servers, without checking out how their email forwarding works. It turns out they have a blacklist you can't opt out of. Very very annoying, to say the least - it's absolutely essential for any thinking being to be able to see what your spam filter lists as spam and correct it on occasion.
Yes, I will change hosting provider. As soon as I get the time. But I would of course prefer GoDaddy changing to being reasonable, and at least allowing me to opt out of having their blacklist applied to all my incoming mail.
I don't see anything wrong with getting paid for mentioning / exhibiting brands (or even arguments) in your art. This is already commonplace in movies, and has been done on occasion by high-profile musicians and novelists.
Given that most people seem to have little against advertising, as long as the ads are good (but typically has a lot against many very bland commercials that litter our mass media), this seems like a win-win situation to me.
...it might not be very useful in itself, but it can fruitfully combined with almost anything. So a combination of maths and something applied is a very good idea. So lots of mathematics with some computer science is a very good idea - probably better than the other way around, from a advanced jobhunting point of view. The specifics of various algorithms and development environments are easy to pick up later, while the a proper mathematical background had better be there from the start.
If you have time over, try to throw in some other random sciences as well - a bit of physics? Some neuroscience? A short course in geology? Broadening your view is always useful.
That you don't want to work with computers is a problem, as that is what you will end up doing in virtually any science or engineering position. So get over it.
Further, I agree with above poster that if you're thinking about getting a PhD you should not wait too long. Getting used to having a proper salary is a major demotivator. So take maximum a year out of university.
The problem you mention, that you don't know exactly how the system will behave under all possible conditions, is a problem we have with all computer programs, especially those that include learning. On the other hand, this is a problem we have with humans as well. The reason (well, one of them) that we let humans drive is that we have done such extensive testing with humans driving, to see under which conditions they can drive safely and under which conditions the "behave unexpectedly" (icy roads, fog, weird intersection layouts, poorly marked turns etc). In the process, tens of thousands of humans drove themselves to death.
So I wouldn't count out self-driving cars until we have done the same amount of testing with them.
Another way we can try to make self-driving cars safer is to use something like Rodney Brooks' Subsumption Architecture, where the controller is structured into layers. The higher layers are responsible for "high-level" behaviour like navigating and planning fuel consumption, while lower layers do simpler things like avoid driving into walls based on sonar information. If a higher layer falls, the lower layers operate just fine without it. The function of the very lowest layer could then be just to stop the car if all the higher layers fail. Given that the layers are separate circuits, it's very unlikely that all will fail simultaneously.
Much of the design can be automated. I'm myself working on using evolutionary algorithms to design car controllers. So far we do quite simple things and only in simulation, but with a good enough simulation it could then be transferred to a real car. Check out this post, including videos.
What about Counter-Strike? Still violent, but you can choose to play the good guys. Well, kind of the good guys.
And I guess there would be some serious money to be had in evolving good good guys... or good bad guys...
Stefano Nolfi's team (responsible for the research in TFA) usually come up with very solid and ingenious research, and the project mentioned is a continuation of a large EU-funded project with interesting results. But it should be remembered that this research is quite far from any actual applications, in large part because people schooled in classical engineering instinctively distrust anything that has been "evolved". Understandable, as there is often no way to tell why the evolved AI is doing what it is doing, just that it is doing it, and no guarantees that it will continue doing it when you need it the most. Sometime in the far future, however, I believe artificial evolution might be the standard way of creating control and intelligence for various purposes.
While computing power is of course good have, I don't think the computing power of individual agents is a major factor hindering development in this type of AI. Our own research (which is similar in spirit to Nolfi's, from TFA) points to the importance of appropriate sensor setups, environments and tasks, and that much can be done with simple neural networks. (On the other hand, if you work in simulation, much processing power might be needed for simulating the environment of the robot.) I recently wrote an article about this, and how computer games might provide the appropriate tasks and environments.
I don't agree. We basically only need to reward the robots appropriately (just like the zombies in "day of the dead") for artificial evolution to create the intelligence for us. "Wants" and "needs" are just words we use to label certain cognitive mechanisms, out of similarity with how we perceive our own thinking. If having "wants" and "needs" leads to better fitness (higher rewards) for the robot, evolution will come up with those things.
I just wrote a post describing the general idea behind this approach to artificial intelligence - check it out!
If you think swarms of little cute robot dogs are scary, have a look at what's going on in our lab over here:
T he%20Flying%20Gridswarm,%20and%20the%20UltraSwarm
http://cswww.essex.ac.uk/staff/owen/research.htm#
The deactivation button might be there... if those who pay for the research want it there.