this is going to change everything.swaption.netand it's going to kick your ass.happy easter everyone!

Thank you - watched it for awhile. I have been reading about neural nets from time to time - if anyone has papers or books to recommend, that would be useful.

the active mode is going to ruin the arrow keys on my keyboard! :Dthx anyway. can you tell a bit more on these agents?

Very good. Here is a book that I read some time ago:Talking Nets: An Oral History of Neural Networks2000"Some of the mythic people responsible for the foundations of modern brain theory and cybernetics, such as Norbert Wiener, Warren McCulloch, and Frank Rosenblatt, appear prominently in the recollections." Will have a look around on SSRN later today.

trackstar, the point is to try to beat them by controlling the blue bloke with your keyboard arrows :)tagoma, outrun, you can find out how I trained them by clicking on the link on the top right for a little blog post on the subject!certainly more fun coding up some AI in JS than trading bond futures and swaptions. best to code'em up to do my day job.

QuoteOriginally posted by: akimontrackstar, the point is to try to beat them by controlling the blue bloke with your keyboard arrows :)tagoma, outrun, you can find out how I trained them by clicking on the link on the top right for a little blog post on the subject!certainly more fun coding up some AI in JS than trading bond futures and swaptions. best to code'em up to do my day job.Yes, I did that while I was watching them.

Interesting! But this approach contains a serious hidden vulnerability. The neural net model assumes there is some kind of invariant physics in the system (as in volley ball or chess) which the neural net might learn to operate effectively within. Yet, in reality, the counterparties in the market create that physics (unlike volley ball or chess) and if they know that there is a neural net on the other side which will learn the physics then they can create one type of physics for the neural net to learn and then change that physics at a later date to take all the neural net's money. What's interesting is that this strategy of manipulation to mislead others can arise spontaneously such as in collections of neural nets or genetic algorithms playing the Prisoners Dilemma game with each other.Of course, if the first neural net knew that it faced a counterparty neural net that was going to fake the physics, then the first neural net could pretend to learn the physics whilst actually waiting for the counterparty to switch physics and the first neural net would be ready for that. Which ever neural net goes deeper on the "I know that you know that I know that you know ......." recursion of social knowledge will win.But here's the final kicker: the collection of counterparties is also a distributed adaptive learning system with each counterparty potentially being a neuron in that broader network. To truly succeed, the first neural net must be larger and deeper than the sum total of all neural nets that it faces.

QuoteOriginally posted by: outrunQuoteOriginally posted by: trackstarThank you - watched it for awhile. I have been reading about neural nets from time to time - if anyone has papers or books to recommend, that would be useful.The game is really nice. It uses a specific type of machine learning: Reinforcement learning, probably Q learning.Learning to win a game by moving about or make trading profits by sending order are of course identical things!the latter is way more difficult since rules of the game change all the time.

QuoteBut here's the final kicker: the collection of counterparties is also a distributed adaptive learning system with each counterparty potentially being a neuron in that broader network. To truly succeed, the first neural net must be larger and deeper than the sum total of all neural nets that it faces.it is only a matter of time, my friend.

QuoteOriginally posted by: outrunQuoteOriginally posted by: ThinkDifferentQuoteOriginally posted by: outrunQuoteOriginally posted by: trackstarThank you - watched it for awhile. I have been reading about neural nets from time to time - if anyone has papers or books to recommend, that would be useful.The game is really nice. It uses a specific type of machine learning: Reinforcement learning, probably Q learning.Learning to win a game by moving about or make trading profits by sending order are of course identical things!the latter is way more difficult since rules of the game change all the time.yes. what I ment is that both are "reinforcement learning", actions and their effect on performance are not direct linked (unlike supervised learning like detecting cats in images where you have a set of train-images and a given boolean target)yes, true. akimon's game and blog are pretty cool, i must say.

QuoteOriginally posted by: akimonQuoteBut here's the final kicker: the collection of counterparties is also a distributed adaptive learning system with each counterparty potentially being a neuron in that broader network. To truly succeed, the first neural net must be larger and deeper than the sum total of all neural nets that it faces.it is only a matter of time, my friend.So true!Of course, this arms race has other effects. As the traders (both natural and artificial) become more sophisticated and complicated in their dynamics, the trading dynamics become more complex and more indistinguishable from noise.The other side effect is that in an arms races, it's the arms makers who profit most. Those who sell the software and the hardware make guaranteed profits.