Can Speed Racer Be An A.I.?

[jyl]

As you may have heard, Deep Mind's AlphaGo, an AI developed by Google has beaten one of the very best players in the world in the board game Go. Deep Mind won a 5 match series against Lee Sedol, 18 time world champion.

Go is a far more complicated game than chess. There are more possible board situations than atoms in the universe. A computer cannot calculate every possible outcome to determine its next move. Humans play Go with a combination of experience, analysis, and intuition. This was believed by many to be something computers could not do. Until now.

Lee Sedol was stunned at his first loss. He said Deep Mind made no mistakes, gave him no opening. By the fourth match he fought back and won. It was 3 to 1. He said he'd never received so many congratulations for winning a Go match. In the final match, Deep Mind made a big mistake early on, then the AI fought back to take the match and win the series 4 to 1.

Article here.

This gives us much to ponder.

One question, perhaps unconsequential but interesting anyway, is whether an AI could win the F1 Drivers Championship.

Here is how I would pose the question.
- The cars are built for human drivers, so require the AI to conform to that interface. It has to fit in the seat, push the pedals, turn the wheel, tap the paddles.
- It has to be self powered for the race duration and weigh as much as a human.
- It can receive only the sensory inputs that a human driver does. Visible light, sound, vibration, acceleration, add air pressure and smell if you wish. No LIDAR, no radar, no infrared, no GPS.
- It can have the same information and learning experiences that a human driver has. Practice drives on the tracks, practice in the same simulators the humans use, actual races against humans or against other versions of itself, starting with lower class cars- even karts, if you like - and working up to F1. No centimeter-scale track mapping uploaded to memory. During the race, information sent by the team to the AI can only be what a team could tell a human driver. Can't send the telemetry data from the car directly to the AI or anything like that.
- I would even consider restricting its raw reaction time and the speed at which it actuates the controls to human levels.

The point is not to see if a machine can beat an F1 car + human driver. The point is to see if an AI "brain" can beat a human brain. As much as possible, we'll make the machine (car), the physical body, the inputs and available data, all equal.

Can the AI, let's call it "Speed Racer", be the WDC?

What do you think? Why do you say that?

Is your answer influenced by whether the AI has a goal of self preservation (aka fear of death)? Should that matter?

[cashflyer]

I think what would matter more is if people will pay to be in the stands to watch racing automatons. May as well throw two blobs of paint on a wall and watch to see which one runs to the bottom first.

[Amail]

I think AI would eventually come out on top. AI can recall and use similar experience faster, plus that whole missing fear of death thing puts it way ahead.

I'd be curious to see how it would respond to human competitors. Humans will race differently against the machine than they would each other (I'd guess more aggressively), so I imagine the machine would race differently in response.

[GH85Carrera]

How big was the computer and power supply needed for the Deep mind? Can it be fit in the space of a human driver?

[stealthn]

I find Go a much easier game than chess

[jyl]

I don't know, about the hardware used. You can fit a hell of a lot of computing power in 160 lb so I'd guess that would be eventually be a manageable constraint.

Yes, humans act differently than computers, but the same challenge was faced with Go. In match 4, when Sedol played unconventionally, Deep Mind was thrown off, but it adapted and came back to win match 5.

I don't think an AI F1 driver would have many fans, but who knows. I don't see why it wouldn't be exciting racing. Deep Mind apparently played unconventional, unusual strategies and by all accounts the matches were exciting, as far as Go matches go :-)

Here is the deeper question, that we're discussing in the context of F1. How good can AIs be at human activities.

We talk sometimes about self driving cars. If an AI will be a better, safer, more reliable driver than most human drivers, even professional drivers, then will there be a role for human drivers?. Maybe the millions of truck drivers, cab drivers, Uber drivers, train conductors, airline pilots, all go away. Or have to work for wages competitive with the ever-declining cost of computing. Even a minimum wage worker costs quite a lot in annual expense, compared with the one-time cost of buying a computer.

Yes, AI software costs billions to develop, but maybe it can be amortized over thousands of applications and tens of millions of instances. Deep Mind is not a specialized AI that can only play Go. It learned and mastered dozens of video games, just by playing them. Suppose you own a trucking fleet and could replace each human driver with a $10K AI? Wouldn't you do it?

"Yeah, that's just low skill work" might be a response. But if an AI can be a world class F1 driver, then can an AI be a top airline pilot? A top fighter pilot? A top neurosurgeon? A top securities trader? A top software developer? A top chef?

Right now, "high skills" are one way that people can win in the struggle that is the economy. If you are a skilled pathologist or radiologist, you make a fine living. But can an AI do as good a job at reading CT scan images and stained tumour samples? Suppose it can. Wouldn't hospitals start to replace their human pathologists with AI pathologists? Especially if the AI can use essentially the same interface as the human, so you don't need a lot of capital investment, don't have to rebuild the physical lab and equipment. Replace your $200K/year human pathologist with a $50K AI?. Why not?. Maybe you'd keep the very best human as supervisors and head of the pathology department. But as the decades go by, where do the new human pathologists come from to replace them?

Substitute other jobs for pathologist or truck driver. Some jobs seem impossible for an AI to do. But until now, the job of "Go grandmaster" seemed impossible for an AI to do. Now it is clear that, if AIs are allowed to compete, future Go grandmasters will be AI, not human.

I think it would be extremely interesting to set Deep Mind or a similar AI to playing a first person combat soldier game like Call of Duty. Perhaps the first thousand games would be disasters for the AI. But it can play ten thousand games, learning from each one. It can go on multiplayer networks and play against humans, a hundred games simultaneously, learning from each one. And then, I believe, when this AI becomes a top player, it can be replicated for just the cost of the computing hardware. Is it a huge step from that to an AI that can actually be a combat soldier?. Okay, that raises mechanical engineering challenges that might be hard to solve. But how hard are the mechanical engineering challenges of fitting an AI into the cockpit of an F18, F35, or whatever the US fighter plane in 2030 is?

[jyl]

Originally Posted by stealthn

I find Go a much easier game than chess

At a low level, of course, since the rules are simple.

At a high level . . . there are more potential Go positions than there are atoms in the universe.

[jyl]

Okay, I looked up the hardware for AlphaGo. It is usually run on pretty hefty computers, 48+ CPUs. The hardware used for this match was 1,920 CPUs and about 280 GPUs.

https://en.m.wikipedia.org/wiki/AlphaGo

So today you couldn't fit AlphaGo (that's the name of the Deep Mind AI software) in an F1 car. But in a decade or two . . .

After all, today's self driving cars don't have racks of supercomputers in the trunk. And presumably the Deep Mind designers aren't (yet) trying to reduce resources consumed.

[flipper35]

Can the bot see and smell the oil on the track? Racing is completely different than chess or Go. Sure, the bot can tell when it starts raining and adapt quicker after the fact but can it anticipate future conditions? Does it understand which drivers are jerks and which ones are clean?

Given your constraints I think it will be a while before AI will beat a human driver.

[GH85Carrera]

I can see the day when a truck that has an AI "driver" is hijacked for the AI parts. Just dump the program and reprogram. Will the AI truck driver be equipped with defensive capabilities?

Will there be a booming market for stolen AI components? Than a set of plans on the internet will allow you to reprogram your stolen parts into your own personal AI needs.

[GH85Carrera]

I did not find and reference to the physical size and power needs of Deep mind.

https://en.wikipedia.org/wiki/Watson_%28computer%29

From the wikipedia article:

Watson is the computer that beat the humans at Jeopardy. It was rather large. And not at all mobile. The system is workload-optimized, integrating massively parallel POWER7 processors and built on IBM's DeepQA technology,which it uses to generate hypotheses, gather massive evidence, and analyze data.[2] Watson employs a cluster of ninety IBM Power 750 servers, each of which uses a 3.5 GHz POWER7 eight core processor, with four threads per core. In total, the system has 2,880 POWER7 processor threads and 16 terabytes of RAM.