This seems to me a very American view of society. Your conclusion would perhaps be correct if the American way would be they only (or only feasible) way. In other countries, lawyers and money can typically buy you much less than in the US.
Now imagine what happens when someone else, e.g. China, starts manufacturing these cars. Imagine what happens when they are used there in large quantities. Imagine what happens when the US press catches wind of the extremely reduced number of accidents in countries that allow these AI cars.
Alternatively, it is not hard to write a law that assigns the liability unambiguously to the person using the car. Then, would you - as a consumer - use an AI taxi cab, if you where possibly liable for any malfunction? Perhaps you'd be urged to say you rather wouldn't. But what if you had the choice between a human driver - with an expected chance of death-by-accident of say 0.001% - and a robot driver - with an expected chance of death-by-accident of say 0.000001% and an expected chance of liability due to a non-lethal accident caused by the robot of say 0.0001%. Which would you choose?
I'm not saying such a law would or should be passed. I just don't share your conclusion that the question of liability will prevent these AI solution from reaching the general public.
Actually, having reread your initial post, I largely agree with you. I think RL should be able to solve Ms. Pacman without the need of higher level strategies that are either hand crafted or learned with the cross-entropy method described in the article. However, the article gives little useful information about this method so I cannot tell whether this can be sufficiently be descripted by the term 'scripting'. I feel sufficiently complex scripts can in fact be called AI. However, this is always a matter of subjective interpretation and valuation of algorithms.
I feel I must comment since I am familiar with the AI used in this case: Reinforcement Learning. RL is a method of finding a mapping of states to actions in a setting where rewards can be obtained. The interesting part is that RL algorithms can learn to behave optimally when only very basic information is given. For instance, it should be enough to simply give small rewards for eating the dots and large punishments for being caught by a ghost. There are many theoretical results in the field that also hold in the case of stochastic environments (such as when the ghosts move randomly). In a sense you don't have control over the learning process, at least not in the sense that you control what exactly happens and which actions get tried. However, in the end theoretically still perfect behavior can be learned. This may take quite some time though, but fortunately good behaviors usually emerge much sooner.
That being said, it is relatively easy to apply these techniques to games such as Ms. Pacman. Much harder problems have already been solved using RL algorithms. What seems missing in the article (though I don't know if this is also the case in the actual research) is comparisons with other RL methods than their own. Though their approach sounds promising and it's nice that they beat some human players, this is not uncommon in games for RL.
Actually, the lower bound has been proven to be at least 11. From the link:
Graham and Rothschild (1971) also provided a lower limit by showing that N* must be at least 6. More recently, Exoo (2003) has shown that N* must be at least 11 and provides experimental evidence suggesting that it is actually even larger.
Re:Only a matter of time
on
The Los Alamos Bug
·
· Score: 2, Interesting
First, define intelligence for me. I dare say there is not much more to our own supposed intelligence than can be accomplished with programming. Of course, I do not mean the rule based AI-like systems used for instance in games and most industrial applications. I mean self organising and/or learning systems powered with algorithms like Reinforcement Learning and/or Neural Networks. I have myself programmed such algorithms to find solutions on tasks I would have never been able to find. Usually these tasks are control based, but hierarchically extending them, you can find solutions for arbitrarily hard problems, including - I believe - the problem of living.
Another question is whether we want to do that. Will it increase overall human happiness?
Anyone care to explain the relevance of a virtual processor for performance issues? Four physical cores means at most four processes running parallel, no?
This is not intended as a sarcastical remark; I'd really like to know...:)
A system like Musicbrainz might work. No complicated rating systems used there, you just suggest a change and then members can vote. From their site:
If 3 people unanimously vote for or against a suggested change, the change is immmediately accepted or rejected, respectively. If the votes are not unanimous the system will wait for 1 week and then accept the suggested change into the database if the simple majority of moderators voted yes. If the vote fails to achieve a simple majority it is dropped from the moderation table without being applied to the database.
Also, they support some automoderations by users that have contributed enough in the past. Of course, information on albums and artists is a lot less contraversial and usually easier verifiable than some of the content on Wikipedia... Still, something similar might be useful.
Slashdot Mirror
This seems to me a very American view of society. Your conclusion would perhaps be correct if the American way would be they only (or only feasible) way. In other countries, lawyers and money can typically buy you much less than in the US.
Now imagine what happens when someone else, e.g. China, starts manufacturing these cars. Imagine what happens when they are used there in large quantities. Imagine what happens when the US press catches wind of the extremely reduced number of accidents in countries that allow these AI cars.
Alternatively, it is not hard to write a law that assigns the liability unambiguously to the person using the car. Then, would you - as a consumer - use an AI taxi cab, if you where possibly liable for any malfunction? Perhaps you'd be urged to say you rather wouldn't. But what if you had the choice between a human driver - with an expected chance of death-by-accident of say 0.001% - and a robot driver - with an expected chance of death-by-accident of say 0.000001% and an expected chance of liability due to a non-lethal accident caused by the robot of say 0.0001%. Which would you choose?
I'm not saying such a law would or should be passed. I just don't share your conclusion that the question of liability will prevent these AI solution from reaching the general public.
And so I did.
Actually, having reread your initial post, I largely agree with you. I think RL should be able to solve Ms. Pacman without the need of higher level strategies that are either hand crafted or learned with the cross-entropy method described in the article. However, the article gives little useful information about this method so I cannot tell whether this can be sufficiently be descripted by the term 'scripting'. I feel sufficiently complex scripts can in fact be called AI. However, this is always a matter of subjective interpretation and valuation of algorithms.
I feel I must comment since I am familiar with the AI used in this case: Reinforcement Learning. RL is a method of finding a mapping of states to actions in a setting where rewards can be obtained. The interesting part is that RL algorithms can learn to behave optimally when only very basic information is given. For instance, it should be enough to simply give small rewards for eating the dots and large punishments for being caught by a ghost. There are many theoretical results in the field that also hold in the case of stochastic environments (such as when the ghosts move randomly). In a sense you don't have control over the learning process, at least not in the sense that you control what exactly happens and which actions get tried. However, in the end theoretically still perfect behavior can be learned. This may take quite some time though, but fortunately good behaviors usually emerge much sooner.
That being said, it is relatively easy to apply these techniques to games such as Ms. Pacman. Much harder problems have already been solved using RL algorithms. What seems missing in the article (though I don't know if this is also the case in the actual research) is comparisons with other RL methods than their own. Though their approach sounds promising and it's nice that they beat some human players, this is not uncommon in games for RL.
Still, it's quite a range.
Wasn't there cocaine in Old Coke?
Another question is whether we want to do that. Will it increase overall human happiness?
This is not intended as a sarcastical remark; I'd really like to know... :)
If 3 people unanimously vote for or against a suggested change, the change is immmediately accepted or rejected, respectively. If the votes are not unanimous the system will wait for 1 week and then accept the suggested change into the database if the simple majority of moderators voted yes. If the vote fails to achieve a simple majority it is dropped from the moderation table without being applied to the database.
Also, they support some automoderations by users that have contributed enough in the past. Of course, information on albums and artists is a lot less contraversial and usually easier verifiable than some of the content on Wikipedia... Still, something similar might be useful.