A Poker-Playing Robot Goes To Work for the Pentagon (wired.com)

In 2017, a poker bot called Libratus made headlines when it roundly defeated four top human players at no-limit Texas Hold 'Em. Now, Libratus' technology is being adapted to take on opponents of a different kind -- in service of the US military.

From a report: Libratus -- Latin for balanced -- was created by researchers from Carnegie Mellon University to test ideas for automated decision making based on game theory. Early last year, the professor who led the project, Tuomas Sandholm, founded a startup called Strategy Robot to adapt his lab's game-playing technology for government use, such as in war games and simulations used to explore military strategy and planning. Late in August, public records show, the company received a two-year contract of up to $10 million with the US Army. It is described as "in support of" a Pentagon agency called the Defense Innovation Unit, created in 2015 to woo Silicon Valley and speed US military adoption of new technology.

[...] Sandholm declines to discuss specifics of Strategy Robot's projects, which include at least one other government contract. He says it can tackle simulations that involve making decisions in a simulated physical space, such as where to place military units. The Defense Innovation Unit declined to comment on the project, and the Army did not respond to requests for comment. Libratus' poker technique suggests Strategy Robot might deliver military personnel some surprising recommendations. Pro players who took on the bot found that it flipped unnervingly between tame and hyperaggressive tactics, all the while relentlessly notching up wins as it calculated paths to victory.

Comments

Shall we play a game?

[alvinrod]

I'm a little uneasy about this. I'm also uncertain as to how you necessarily go about training this. With poker it's not particularly hard to find real opponents who will legitimately play their best. With this it seems like all you really do is train to to be really good at beating your own army.

Re:Shall we play a game?

[alvinrod]

The problem with training a bot against other bots is that it might not be good against human players. With a poker bot, you can always find some human players to pit it against, maybe even some professionals who think they can win some easy money.

You can't do the same quite as easily with a military bot though. Sure you can have it play against humans in field exercises, but you're having it play against your own army. In a certain way you're actually creating the best possible bot your enemy could want. You can try to get around this by attempting to emulate your enemies tactics, but that's always imprecise and surely something that the enemy already expects.

Re:Shall we play a game?

[AHuxley]

Re "With this it seems like all you really do is train to to be really good at beating your own army."

Consider the politics with the US Army, Navy, NSA, CIA, special forces, other US intelligence contractors.
The math smarts and skills needed, the level of fitness.
Political changes to the demographics, fitness, quotas, virtue signalling.
The ability to attract people with smarts and sports skills. To make the US intelligence community take on a different demographics.
The IQ level and the trust placed in middle and upper levels of the US mil vs the rapid political changes to US intelligence.
Then mix it up with political considerations. How their side of US politics is doing.
Does a general, the top level of an agency, law enfacement, talk to the US media about their support for one side of US politics?
Walk out with secret documents to give to the US media to make a domestic party political statement?

The US Army and Navy might see such partisan political problems as their time to demand power and budgets back from partisan agency leadership.
The Army and Navy once did crypto, spying and intelligence to a better standard than todays "agency" contractors. Why not revert to a better more secure leadership?

The buddy system, lie tests, FBI investigations don't find the political motivated staff in time and they walk out with decades of US secrets and methods.

What if an AI robot network could make a lot of the US mil work better?

A party political appointee to an agency over decades would have less access and less ability to talk to the media about things they no longer have access to.
The robot AI warns the US Army of a next generation of "Tet Offensive" on time rather than what the political feelings are of US intelligence agents to US politics over that decade.
The AI robot reports to the US Army and Navy every time that the enemy is changing, on the move, activating their spy network in the open..
The US intelligence agents may be thinking of the optics of politics and their side of politics. Warnings never get passed on as its not their side of politics.
The robot AI brings raw intelligence back to the US mil and away from the political optics of other US agencies.

Re: Shall we play a game?

[ememisya]

Even then this thing is probably just "doing its best". Think about it this way, if you think it's better to have this installed in a critical meeting room than not for reasons of security, and that is exactly what will happen, people will come prepared. Such systems don't know if you're lying, they look at the physiological changes and historical patterns in response to your actions. Okay, let's say if somebody asks you is your name Steve, and it says so on your ID, you might reason, "I don't know? What if I'm named Ron who is in a hospital in a coma for 40 years." Boom, system just showed patterns of hesitation and creative thinking, clearly this person is lying and his name is not John. Peoplr will come prepared and genuinely panick when they have 4 of a kind aces because some maniac called 4Aces hacked their credit cards, maxed them out, insurance didn't cover it all and now they are scared. The machine will grin and go all in, "He's bluffing, he's scared." Aw fuck.

All hail President Libratus

[Kryptonut]

All hail President Libratus

Re:All hail President Libratus

[Kryptonut]

Hence why this robot will take over as President

Re:All hail President Libratus

[Opportunist]

Well, he's pretty used to playing va banque...

should not Bezos always win in poker?

[mapkinase]

Obviously I have very little idea of how poker works.

Professor Falken was unavailable for comment

[Etcetera]

Sometimes I feel like the generation in power in both Silicon Valley and Washington, DC have forgotten (or never watched enough) dystopian 1970s and early '80s sci-fi.

SkyNet, VIKI, MCP, Colossus had to start somewhere

[schwit1]

Poker seems as good a place as any. But I hope we've learned not to give it the keys to the nukes.

Bad summary

[Blue23]

Sorry, but that's a misleading summary for technical news. Libratus did some pretty good playing, but saying it beat four top human opponents is extremely misleading.

What it did do was play thousands of rounds one on one. With exceedingly large bankrolls compared to the size of the big blind that were reset after every hand. In other words, it never had to play with short stack, never had to worry that the opponent couldn't cover it's own bets, and that really long shots (which are easier for a computer to calculate) can be made to pay off if hit because of the size of the bankrolls were much larger than usual for the size bets being made. And was only one on one, so it had a minimum of unknown information, betting and bluffing. Hold 'em, so 5 common cars and only two hold cards it doesn't know. And thousands of rounds each, so any small edge would have time to multiply.

Now, it did do this against four top players (each against their own copy of Libratus). It really was quite an accomplishment. But it's not nearing the general poker imperfect-information feint-analyzing multiple-unknowns that the summary makes it out to be. Come on /., be News for Nerds. Get the tech details right.

Re:Bad summary

[Shotgun]

So, the computer is able to win, because it is optimized for counting cards, with "bluffing ratio" being one of the cards.

If war strategy is about being able to keep track of thousands of small details and their relative impacts, AI like this would be useful.

Re: Bad summary

[tdelbruck]

Totally agree, this summary really gets the limitations much better. Not at all clear how heads-up poker translates to war games. They must be working on multi player tourney now?

Re:Bad summary

[Blue23]

You must watch a lot of movies. In a real poker game the chips on a table are already paid for, you don't have to worry about whether you or your opponent can cover a bet they make and any serious professional player will never play with money they can't afford to lose in that session. Most cash games are generally deep not short stacked. Heads up is a very common cash game format online, it't not some rare contrived idea that doesn't have a real world equivalent. Being a successful poker player isn't about looking for twitches in your opponents face, rather just a string of mathematical optimization calculations.
As an example: if I am faced with a bluff on the river my decision is not concerned about trying to figure out if my opponent is bluffing in this specific hand as I can't answer that due to their cards being unknown to me. My decision is based around working out approximately what % of the time a correct strategy would bluff in that instance, what the correct call / fold ratio is with a bluff catcher to that bluffing frequency and if I have noticed my opponent bluffing too much / not often enough so I can adjust my ratio to exploit them. That's precisely what this bot does except the bot does it far more accurately.

I love watching movies with poker, they are so funny. Or watching most of the poker shows on TV when that was big, since they just show the exciting hands and give casual players the wrong idea about how to play.

I don't have your claimed chops - I never supported myself via poker. But I think you're making some bad assumptions there. Let's check what I was saying that the /. summary was poor.

Libratus only played heads up. The summary implied it was playing five handed. These aren't the same.

It played 120K hands, and it never had to deal with the repercussions of a hand because it was reset. It's like if you, playing poker, never had a bad beat that put you at a disadvantage. And never had to worry about covering an all in - or if your opponent had enough to cover yours and make something worthwhile. This also isn't regular poker.

Basically, it was set up to minimize extraneous inputs (multiple players interacting with each other), minimize unknowns, and could exploit narrow margins over a long period of time (120K hands) without ever having to worry about making a bad play and losing their stake.

What I said is that the summary was bad in presenting it as a general poker playing savant, and I stand by that.

Re: Bad summary

[mimi210]

With poker it's not particularly hard to find real opponents who will legitimately play their best. With this it seems like all you really do is train to to be really good at beating your own army.

Re: Bad summary

[Blue23]

Just to add I'm still not sure what you mean by "covering an all in". No limit hold em should really be table stakes hold em. The maximum you can lose in a hand is the amount you had in front of you at the start of a hand. If you start with 1000 and I start with 200 and you shove all in, I call with the 200 I'm playing with. I don't have to cover the full 1k to call.

As for heads up v 5 max, yeah, totally different games.

Implied pot odds.

Pot odds can normally have you stay in when you have less than a 50% chance to win, that's joe standard play. But there are times when you have a long shot to get the nuts where it's not just what's in the pot, but what more you expect you can get out of them. If they started then hand with 200 and only have 20 left, there's a lot less potential upside then if they started the hand with 1000 and have 820 left.

Not that you should call only if they go in for 820, but there are times that if you catch their card and they will have gone in for 120-150 more then it's worth it, but wouldn't be for 20 more then what's in the pot.

what a great time to be alive

[sad_]

we get both skynet & wopr irl

Re: what a great time to be alive

[mimi210]

With poker it's not particularly hard to find real opponents who will legitimately play their best. With this it seems like all you really do is train to to be really good at beating your own army. [url=https://audacity.onl/]Audacity[/url] [url=https://findmyiphone.onl/]Find My iPhone[/url] [url=https://origin.onl/]Origin[/url]

Possibility to refuse orders vital

[Roger W Moore]

Sorry but the picture you paint is a nightmare waiting to happen. Making an army entirely beholden to one person's will without anyone anywhere being able to delay, question or even subvert their orders if they give illegal ones is a disaster waiting to happen. It may only be required in exceptionally rare circumstances but a human knows that, under those circumstances, they can almost certainly get away with disobedience (or may just be willing to suffer the severe consequences of disobedience regardless) whereas a computer may not.

Knowing that a general may well refuse to follow an illegal order and may publicise the order s/he was given also helps prevent politicians from giving those orders in the first place. Having everything under the absolute control of one person is a well-known recipe for disaster. By all means have AI providing advice but unless we have people, with all the flaws and problems you mention above, in command, we will have a disaster far worse than the problems you are trying to fix.

Re:Possibility to refuse orders vital

[AHuxley]

The problem is the political problems within the US intelligence community cant be fixed.
Generations only want to support their side of US politics.
The quality of raw intelligence they hold back and pass on depends on their political views.
Too many split loyalties, people of another nations faith, people of another nations politics.
Are totally supporting another faiths "freedom fighters" and don't mind using up the US mil to support the "freedom fighters".
Its then the Army and Navy who gets sent on failed missions by such creative advice.

Hand back such work to the Army. Try an AI to sort the raw material for the Army and Navy.

Re:Possibility to refuse orders vital

[Roger W Moore]

If you really have people who have split loyalties then you have to fix that issue first. If you try to avoid it by centralizing all authority in one person then what happens when that one person has split loyalties and puts political ideology etc. over duty to their country? They'll have nobody else with conflicting views to temper their power. Information technology amplifies what a single person can do - if you have a problem with what people are doing you need to fix that first before you amplify their capabilities!

Re: WOPR

[ememisya]

Oh look! Another lie detector.