Yeah my bad. I thought that it still had to be FE related.
I don’t think the Google AI dominated go, but it certainly won against the world’s top player. Another AI, not developed by Google, defeated a bunch of human’s in a poker game. What made this interesting is that unlike chess or go, poker has a component of hidden information and theory of mind components in betting. The AI adapted to both of these elements and learned to call bluffs.
I've actually written poker algorithms before. I wrote one for texas holdem based on the probability of winning any hand based on what's in my hand, whats on the table and what others could conceivably be holding by probability, what I might draw on the river, what they might get etc (odds for each stage). The problem is, whilst it worked and could say "you have a 45% chance of winning this hand over the other 4 people left in", I could never wrap my head around the maths of how much to bid to lure maximum money from people, to stop them folding because my algoritmn was predicting say 97% chance of winning so it went all in etc. I could get it to work out if the pot money vs the next stake was worth it to 'see' my opponent cards, but should my algorithm raise by $4, $6, $7? I had no idea how to sort that. And it seemed how much you bid, is actually more important than the cards you hold ... you hold average cards, you need to stay in but not lose too much.
I have recently been looking at tensor flow (from a hobbiest point of view), and it might be able to solve these problems for me, and as you say, be able to call a bluff ... which I wasn't even close to getting the maths right for. I could only say, odds in my favour, bet, not in my favour, don't bet ... binary ... and that won't beat a top player even if I know the odds. I did have 'the gun' in my probability, and the algorithm I wrote would know where on the table it was and calculate the odds dependant on its seat when asked to bid ... an instantaneous set of odds.
But I won't be using tensorflow for this ... gambling sites are already onto this and now actively hunt down signs of machine learning. That window has passed.
Machine learning would be the ultimate answer to earth's shape. Not even Tom Bishop would argue because it is based on observable science. You don't give ML any assumptions. You just feed it data and it iterates repeatedly until it finds the answer. The problem with ML in today's format, is that whilst we'd end up knowing what shape the earth is, we'd have no idea how the machine came to that conclusion, we'd only know it is right. Much like we have no idea how Google's AI plays chess. It just does it.
I have a theory that ML will actually cause a 'great ignorance'. Lets say you got ML to start predicting the weather. Now if would just look at all the data from ocean buoys, airport reports, temp, pressure, visibility, dew point etc etc. And it would work out the weather ... and it would be far more accurate than anything we have today. Maybe we'd end up with a 30 day forecast. People at the weather service would abandon trying to predict the weather, the machine does it better, but no one knows how it does it. So, you'd have a meteorological office filled with people that could write machine learning, and no one who actually knew how to predict weather themselves from data as no one is employed to do that. The science would grind to a halt. There is no point learning a solved problem, as it is useless, weather isn't predicted that way any more.
Span this through multiple industries such as medical cures, accountancy, logistics, etc ... no one would have any knowledge or skill whatsoever. But I guess this is why people think AI will kill jobs for billions of people. We'll all be dumb, unskilled and unneeded.
Show Posts
