Slashdot Mirror

Yes, years. Libratus uses 16 Terabytes of memory for just a 2 player game. The size of the game tree increases by at least a factor of 1000 when moving up to just 3 players.

This is why better algorithms are almost always a bigger factor than increased computing power when solving these problems. They won't solve playing against 3+ players with more RAM, they will solve it with better algorithms. By some cases algorithmic improvements can be 43 times more important than computing power improvements.

Considering this AI is already dealing with unknown information, I doubt the size of the "game tree" increases by the factor you cited as you add more players.

John Cook (put his blog in your RSS feed if you don't already have it) made a very good point recently: The speed gains from Moore's Law are dwarfed by the speed gains from algorithmic improvements. And unlike Moore's Law, we're not yet seeing a limit approaching for better ways to solve stuff. The post in question: http://www.johndcook.com/blog/...

The best comment I have seen about Python and R:

I prefer Python to R for mathematical computing because mathematical computing doesn't exist in a vacuum; there's always other stuff to do. I find doing mathematical programming in a general-purpose language is easier than doing general-purpose programming in a mathematical language.

http://www.johndcook.com/blog/2012/10/24/python-for-data-analysis/

genius programmers thought it was great to blindly map color to grayscale

Huh, what do you mean? g = (r + g + b) / 3;

That's probably exactly what they did (or what the display driver does as the browser probably doesn't adjust the colors at all).

But when doing it this way, you often get bad contrast (slightly-darker-gray on slightly-lighter-gray). See this page for an illustration with pictures.

Apparently Twitter supports Unicode, which opens a whole range of new possibilities ;-)

http://www.johndcook.com/blog/2009/01/14/how-to-insert-graphics-in-twitter-messages/

True, but any engineer will tell you that you learn far more from your failures than your successes.

If one looks at the fatal accidents of the Shuttle program, an odd aspect becomes apparent. In both cases, the failure mode was observed in prior launches. For example, Feynman notes that considerable burn through on the o rings was accepted prior to the Challenger accident. Similarly, ice and foam strikes were common and looked for at each launch (the foam strike that doomed Columbia was seen by ground video within hours of launch). In each case, the cause was known ahead of the actual accident (an engineer warned of fragile o rings at freezing temperatures prior to the launch of Challenger and there was a debate about the foam strike on Columbia for almost its entire stay in space).

The problem in each case was human error on the ground. The deciders chose to ignore the warnings from previous launches as well as the warnings of engineers for the current launch.

But my point here is that while one successful launch is not as good a learning experience as one failure, a lot of successful launches may well be more valuable. With the Shuttle, each of the failures had been preceded by close calls from previous launches.

For a final example, during the Second World War, the UK military was trying to figure out where to armor their bombers. This is the exact inverse of your engineering problem in that the failures never made it back and for the most part, could not be studied. They could only study the successes. Statistician Abraham Wald figured out that the bombers should be armored exactly where the returning bombers had never been shot.

I think in a large enough program, you could deliberately break rockets and components. For example, launch some first stages with deliberate flaws. That might be useful failure testing (at least it tests your telemetry collecting ability).

The pressure difference theory accounts for only a fraction of the generated lift. The majority comes from the the reaction from deflecting the air downwards.

Laminar flow causes the air to stick to the top of the wing and is redirected slightly downwards, the underside pushes yet more air downwards in a more obvious way. The vertical component of this is what generates most of the lift.

Couple of quick links.