Hacker News Reader

Some rules can be derived from instrumenting humans as they perform the maneuvers, and generalizing from their behavior? We used to instrument motorcycle riders at Harley Davidson and create fuzzy-logic models of expert riders as they performed certain acts on a track (dodging road hazard; emergency stop; hairpin turn). Our goal was also a fuzzy-logic driver model, which they used to help design new motorcycle suspensions/steering that would feel 'natural' to an expert rider e.g. mesh well with the model they had for an expert.

I asked myself the same question, and did read some papers, but could not find a recent comprehensive survey on automatic fuzzy rule generation (I admit I gave up after ~15 minutes).

What I found did not convince me that it would fare better than an off-the-shelf (somewhat) non-interpretable statistical supervised learning algorithm.

It can be a nice way of bootstrapping the rule writing process, or to go the other way : to discover and analyze new expert knowledge by looking at the rules.

But performance-wise, I would go the machine learning way anytime.

Also, Inverse Reinforcement Learning seems to be very promising : one guesses the reward function by observing the expert acting.

I imagine there's probably significant regulatory constraints into the interpretability of any models generated to run combat weapons platforms, even if just in a simulation. Deviations from a norm during times of war due to needing to chase after additional data or testing hypotheses might be considered a significant demerit to the model. Alternatively, formalizing these rules may be helpful for instructing new pilots or adhering to existing rules-of-engagement.

Quinlans https://en.wikipedia.org/wiki/C4.5_algorithm is somewhat popular for similar tasks, it allows to build decision trees from data that are conceptually similar to such fuzzy rules, and the rules can be human-readable so they can be really powerful after expert review.

For example, a very specific condition can either mean that this particular condition is useful, or that simply the training data happened to have those particular instances of a more general condition - and a human expert can usually easily decide which rules need to be extended for proper generalization beyond your training data, but the automated generation helps identify factors that the expert could recognize, but wouldn't think of if doing it themselves from scratch.

Thanks for the ref !

I'd like to double-click on that comment. :) If you're ever interested in writing more about that, I'd upvote it.

Yeah it was weird working with guys named 'Roadkill' and 'Slash'. Nice as could be. Terrific riders.

Wow that's very innovative for HD. Why don't they put some of that innovation effort into their actual drivetrain? I loved my Buell's look and ride, but I mistrusted it's horrid 50-year-old Baker transmission which went on to completely fail at ~6,000 miles, necessitating me to disassemble the entire engine to split the crankcase so I could repair it. After seeing its guts, I no longer wanted it. It ignores a half century of innovation in motorcycle design producing a machine that I vote most likely to unexpectedly leave me on the side of the road.

I'll throw out a guess that part of it is the character of the bike being tied to the drivetrain.

Even outside Buell, which was a bit of a neither-fish-nor-foul anomaly, they've had the odd innovative model here and there--the V-Rod comes to mind. And I've seen some interesting things in their ABS systems and some other components. But I think the most successful models have been very conservative about their drivetrain as it's part of their signature sound/feel.

My hope is that Polaris' recent critical success with the Indian Scout (which I bought over HD--far better bang for buck than a Sportster) gives the segment a kick in the ass.

The continued anomaly of the Buell is partly the V-Rod's fault. The V-Rod engine was originally supposed to be for the Buell line to address the issue until the mothership got interested in it. They added too much weight and too high a deck height for Buell chassis, so the project was wrestled away from Buell to make a bike that, in the end, HD couldn't really sell anyway.

Interesting--I had no idea that the V-Rod engine was originally destined for Buell. That makes lots of sense.

Buell had some great innovative designs too, but you could tell that was all Erik and not the motor company. I wasn't surprised when the split happened.

What many don't know about the shuttering of Buell is that it happened while HD was in the process of receiving over a billion in TARP money. This information was kept secret for over a year. Harley had something like ~6000 employees at the time. Buell operated on ~100. It seemed like a ludicrous move; like it had to be political.

If you're looking for innovative motorcycling, gotta look at a sportbike. HDs are for people who love cruisers, a genre defined by rat bikes and self-made bikes, but for people with money. (Not knocking on rat bikes, though, those things are awesome)

what, what is fuzzy logic model? From parent post, it seems to be a like data learning, but manually?

It's an approach to AI that allows you to generate rules based on probabilistic logic (0.0-1.0) rather than strictly boolean true/false (1 or 0). The applications are used in tons of different systems from medical diagnostics tools to washing machines.

The plus side is that it allows you to make systems that can be tweaked using trial and error to handle cases that would require very complex logic otherwise. The down side is that alot of the time the systems aren't provable the way other types of logic are, and they can be a pain to debug.

I would assume that in a military setting (as it is in most bureocratic/management settings) a solution like this has the immense advantage that one can precisely determine the source of any one error to a specific requirement/rule.

It is very hard to ask for management to trust a system they know nothing about, where they have literally no control over final behaviour, even if in the end it will perform better overall. In a rule-based system, instead, it is always possible to make adjustments and blame mistakes very efficiently to specific causes.

I guess this is the main reason why "true" AI is currently being used mostly in information fields, rather than on physical machines and engineering. No-one would know how to deal with the outcome of a fuzzy learned algorithm making the wrong decision. This is also a reason why autonomous cars are very interesting to me, even though I bet they are still full of ad-hoc rules in order to have a layer of "manageability" over the overall system.

I see how it can sound appealing to a bureaucrat, but as a programmer, debugging the concurrent evaluation of thousands of "natural" language IF...THEN... rules until I find the questionable one where a threshold was defined too low or too high sounds like a nightmare.

I imagine they would log all the inputs, as well as branches taken so they can later replay everything in a debugger. That would make the process much simpler.

Being able to take the blackbox recordings of your combat drone which got shot down, reconstruct the scenario and permute the rules till you get a win, seems like...well, a big win.

Air combat is also one of those areas which does notionally have narrowly computable victory parameters - given hardware of capabilities X, there is a model we don't know which should generally predict the outcome.

At least you can actually debug it.

Only if you do it wrong?

If it ends up like thousands upon thousands of if statements then you got it wrong (or so I think).

Getting it right IMO means realizing that rules are not necessarily source code but often data to be iterated over, filtered etc.

This is a form of metaprogramming that can be practiced even in Java ;-)

That would indeed be a nightmare, but trying to debug a machine-generated neural network is actually worse.

Rule-based systems have had for a long time the ability to automatically answer queries about why was such or such decision taken.

Comparing comments here to the comments on, say, first Alpha Go post, reveals the amazing amount of AI bias on this website.

When an expert system beats some human in a complex real-life problem the comments are about how it is narrow, boring, not sufficiently tested and ultimately doesn't matter.

When a neural network (with the help of MCTS and an entire data-center full of servers) beats some human in a board game the comments here hype it through the roof and jump to conclusions about the coming dawn AGI.

I'd guess there's a difference in kind that people get excited about.

The fighter jet AI technique is hard-coded to a very specific problem domain, and could only be reproduced in a different problem domain by doing it from scratch.

The technique used by AlphaGo is at least closer to the idea that we can eventually build generically trainable machines that can learn to do a variety of tasks, without having to code them from scratch every time.

>> The fighter jet AI technique is hard-coded to a very specific problem domain

So are machine learning models, ultimately. Although the algorithms that build the models have more general application, once you train a model on a certain set of data, that model can only ever be used in the domain circumscribed by the data. They're one-trick ponies, yes?

>> we can eventually build generically trainable machines that can learn to do a variety of tasks, without having to code them from scratch every time.

We would however have to train them from scratch every time, for each separate task, and each time we'd need terabytes of data and megawatts of power.

You realize that this system uses machine learning? And that AlphaGo was custom-built to play Go?

AlphaGo has several handcoded training features, uses Monte-Carlo tree search, and was primed by a huge human-made dataset. It's not like it can be throws at other problems without heave re-engineering.

>> Comparing comments here to the comments on, say, first Alpha Go post, reveals the amazing amount of AI bias on this website.

I think it's because only recently has AI been in the news again and it's been the news thanks to machine learning and neural networks in particular (and more specifically, deep learning). The last time there was a big todo about AI was in the '90s and most people writing here were probably not old enough to figure out what the hell was up back then.

Er, I'm not dating myself here. I know about those things out of happy accident (maybe a story for another time). Most people who graduated from CS in the last five or six years will not have heard anything about expert systems and GOFAI, except that it failed etc, if that.

Then along comes Google and promises it can make your phone talk to you. People are intrigued.

But of course, those who don't know their history are doomed to repeat it.

For me personally, this is less exciting because AI jet pilots have a natural advantage.

Human pilots are limited in maneuvering by the amount of force their body can take. Computer pilots are limited by the amount of force the airframe can take. The later allows much more aggressive and radical maneuvers.

That means, compared to playing Go, the robot pilot can be comparatively much worse than the human and still win decisively by taking advantage of high g maneuvers.

The test were done in a simulation. Also, if you want to run high-G maneuvers, you can always control a drone from the ground.

I am not very comfortable with a machine that's very competent in killing fighter pilots. I am much less comfortable with such machines generalizing that competency to other, closer, problem spaces. Also, in cases where the use of deadly force happens without a human in the loop, being able to describe exactly which rules triggered and caused the death of a friendly pilot or that C-40 that happened to actually be a 737 full of passengers would be a requirement. "Because the plane got confused" is not very satisfactory.

How about a future where AI fighter pilots fight against other AI fighter pilots?

Maybe one day war will be less about killing people, and more of a battle between countries' best engineers.

Maybe I'm just optimistic, but I think robot wars would be a hell of a lot better than real wars.

Both sides' civilians could meet together on bleachers and share popcorn.

We are already partially there only with athletes instead of engineers.

(That said I'm less optimistic I guess.)

England may as well just roll over and capitulate to everyone in that case - Go Iceland ;-)

This reminds me of Philip K. Dick's "The Second Variety" that I recently read. (Not that I'm trying to make a prescient political-statement, just sharing a fun short-story you might be interested in.)

http://manybooks.net/pages/dickp3203232032/0.html

Such a great story. Thank you for posting it.

Why are their targets other AI fighter pilots? That's a very, very big assumption. Can we safely assume that the AI pilot won't target buildings/ships/etc.? Even if we assume a well-behaved AI, why should we assume that AI fighter pilots will target military targets? Can we say for certain that they won't be controlled by a malicious despot?

What when one AI pilot generalizes it's knowledge on enemy airplanes and finds out that bombing an airplane factory destroys multiple targets with minimum risk to itself?

>Maybe one day war will be less about killing people, and more of a battle between countries' best engineers.

When two robots fight each other, it's not because the other robot is the target, it's because it's an obstacle between them and the human targets that will actually affect the war.

Because in real wars the body count and the physical destruction is what matters.

Not necessarily. Generally, a conventional war ends when one side surrenders, usually because it knows it can't win and wants to minimize further losses. If we reach a point where unmanned weapons are the ultimate tool for destruction, then it makes sense that by the time the AI combatants reach a point where they could target enemy populations, the war would already be decided, and the nation about to have its people killed would surrender. There will be exceptions, but it would be no worse than current war, and at least potentially be better.

That is of course assuming that the most efficient strategy to achieve a surrender doesn't actually end up being a bypass of enemy AI and direct targeting of human populations. This would only be possible if there's a huge mismatch between offensive and defensive capability. But then you end up with a MAD situation similar to what we have with nukes.

That said, I expect that the flip-side is that this tech would ultimately make a guerilla war by a local population against an occupying force very very ugly for the guerillas.

That would essentially mean two nations agreeing to resolve their intractable diplomatic differences with trial-by-combat, and to abide by the results. It's a lovely notion, but in practice you can't really force a nation to change policy without a palpable threat to their population and/or infrastructure.

So it boils down not to robots killing robots, but robots killing humans, while other robots try to stop them.

Fighter jet AIs aren't that scary compared to the work being done on algorithms for teams of robot soldiers. The US's TARDEC and the Australian DSTO held a competition for this back in 2010 called Multi Autonomous Ground-robotic International Challenge(MAGIC)[0]. In this competition, a team of aerial and ground robots had to perform a simulated combat mission to 'secure' a set of moving(other soldiers) and stationary targets(IEDs) in an approximation of an urban environment.

In simulation, the algorithms demonstrated for doing this do quite well with, being able to complete the mission with a success rate of 97.5%, so long as one has 6 search robots and 3 gun robots.

This did not work so well in real life, partially because real robots are difficult to do with. It is still disturbing though because of the high success rate, not to mention the immediate applicability to robot SWAT teams. As a civilian, I'd be much more concerned with a SWAT AI than a fighter jet AI.

However, robot swat teams are still a ways off.

[0]http://singularityhub.com/2010/03/19/teams-of-military-robot... [1]https://en.wikipedia.org/wiki/Multi_Autonomous_Ground-roboti... [2]http://www.frc.ri.cmu.edu/~ssingh/Sanjiv_Singh/PUBS_CONF_fil...

>The mathematical model of Fuzzy Trees is nice, but this is completely ad-hoc to the specific modelization of the problem, and will fail to generalize to any other problem space.

Well, why should it? No one is inventing HAL-like AI anytime soon, or ever. If this system does a better job of killing the enemy than human pilots then its quite the breakthrough. Projecting air power is one of the ways countries keep aggressors away and this would be quite an advantage for variety of reasons. Not the least of which means you can now design AI driven fighters that have zero design compromises to keep human pilots alive.

I imagine fighter engagement consists of a fairly limited set of problems to solve. Think of this as just a souped up autopilot/autoland system, except with guns and missiles. We're not asking the AI to write the next Romeo and Juliet here.

>because the paper is very light on details

Defense contractors aren't known for sharing details. I imagine this is a competitive advantage and they want to keep their cards close to their chest. There may even be national security issues here.

> Well, why should it?

Because what is trumpeted as a breakthrough may in fact be so narrow in scope that it may not even be possible to use it in a flight combat video game without a lot of work, let alone any real life environment.

It is absolutely, very closely tied to the mathematical model of aerial combat that they devised and can not easily be made to accommodate new insights, or new challenges.

> Because what is trumpeted as a breakthrough may in fact be so narrow in scope that it may not even be possible to use it in a flight combat video game without a lot of work, let alone any real life environment.

What? They were quite literally running the AI in a simulator -- ie, a very expensive video game. The only thing that might not scale is the computational power necessary to execute.

Sorry, I was unclear, I meant in /another/ video game. They are strongly tied with their particular modelling of the problem.

I suppose it depends on the granularity of the rules. Deciding what to do is separate from how to do it. If those are tied together within the rule system, then yes I agree, the rules would need to be tweaked for every new model. But if the rules only define the goal of the system, and flight mechanics are handled by, say, a goal-based feedback mechanism, then it should work fine in any model.

Additionally, doesn't this mean that the code is very sensitive. If the opponent were to get their hands on the code it would be easy to predict what these things are going to do.

So, while it might be the fashionable thing to do some kind of (machine/deep/?) learning approach where you allow it to run millions of simulations and figure out things itself, I can understand why they didn't.

Learning approaches which depend on mass-simulation are great when your problem only ever exists in a "virtual" context, but what happens when you want to take your trained neural network out into the real world? Clearly it's going to have to adapt to the differences between the real world and the virtual world - but how would you do that? You can't run millions of dogfights in the real world to adjust its training.

?

The solution would be to make the simulator so good that there is no practical difference.

It is not about fashion, it is about not being ad-hoc.

For small scale problem where most of the variables are well understood, this kind of approaches work beautifully. Big problem are better tackled by a more generic approach (maybe with some ad-hoc adaptations, such as mixed approach between expert systems and statistical algorithms, feature engineering, etc.) because these approaches will be more resilient to an exposure to the real world, and the manpower invested in them is useful in more than one problem domain.

To address your last point, there is an extensive body of work on data-scarce environments. I've even seen a talk about applying reinforcement learning to endangered species preservation, where you only get a single digit number of interaction with the system !

This is called domain adaptation and transfer learning in the literature. There are ways to do that. It is an active area of research. Basically the idea is to run a few real world dogfights (you could conceivably collect a few hundreds), and use methods to adapt the simulation model to the new domain. Solutions involving unsupervised learning (e.g. no dogfight, just collect sensor data from fighters - you could collect thousands of hours this way) are also active areas of research.

For a start, if it can be done by a fuzzy decision tree, it can be derived by a Bayesian network (that is basically a fuzzy decision tree that keeps extra data for learning) and made more versatile after that.

But the decision tree is much more tractable, thus the longer it's kept on this format, the more future-proof is the work.

Sounds like how Deep Blue defeated Kasparov. The first time is always awkward, but now that we know that it can be done we can develop more generic algorithms. A Stockfish for air combat may be several years away but it's coming.

I'm not sure that an open source AI for air combat would get you very far...depending on the licensing terms.

That's fuzzy logic for you. Probably why it mostly died around the turn of the century.

It's still used in a few systems as a complementary system involving PIC controls.