Forums  > Art  > survival of the fittest  
     
Page 1 of 1
Display using:  

akimon


Total Posts: 566
Joined: Dec 2004
 
Posted: 2015-04-25 07:12


my latest creation.

polysena


Total Posts: 1046
Joined: Nov 2007
 
Posted: 2015-04-25 09:26
pretty cool, and nice, do you program the whole thing?

Свобода - это то, что у меня внутри. (Ленинград и Кипелов - "Свобода")

akimon


Total Posts: 566
Joined: Dec 2004
 
Posted: 2015-04-25 10:02
yeah, it's written entirely using js with a helper libraries

mmport80


Total Posts: 85
Joined: Jul 2010
 
Posted: 2015-04-27 11:03
V nice!

--- http://johnorford.blogspot.com http://blog.johnorford.com

NeroTulip


Total Posts: 997
Joined: May 2004
 
Posted: 2015-04-27 11:58
That's pretty cool, now can you get them to reproduce? I really want to see these things copulate while avoiding planks LOL.

Inflatable trader

akimon


Total Posts: 566
Joined: Dec 2004
 
Posted: 2015-04-27 15:23
Thx!

You can see these things evolve from originally not knowing anything, here. The last 30 survivors of each generation cross-mutate their genes, which is just a vector of the weights of their neural network, to the remaining 70, to generate the next generation of agents.

After a few hundred generations they become half decent.

unsmt


Total Posts: 196
Joined: Jul 2014
 
Posted: 2015-04-27 17:13
it looks that you can imitate a branch stochastic process and changes dynamically rates of birth and death it seems can be sold.

Nonius
Founding Member
Nonius Unbound
Total Posts: 12702
Joined: Mar 2004
 
Posted: 2015-04-29 00:30
Very cool akimon!

Chiral is Tyler Durden

il_vitorio


Total Posts: 103
Joined: Aug 2014
 
Posted: 2015-04-30 21:05
Wow pretty excelent!

One of my most productive days was throwing away 1000 lines of code.

NeroTulip


Total Posts: 997
Joined: May 2004
 
Posted: 2015-05-01 15:36
I can't figure out why nobody has succeeded in feeding something like that with all available market data in order to make a superhuman trader... We got self driving cars, but no super trader... We finance types are retarded, or at least 20 years behind.

Inflatable trader

pj


Total Posts: 3343
Joined: Jun 2004
 
Posted: 2015-05-01 15:46
Inhuman traders are done already

OFFENDERS WILL BE TERMINATED

pj


Total Posts: 3343
Joined: Jun 2004
 
Posted: 2015-05-01 15:47
double post sorry

OFFENDERS WILL BE TERMINATED

curvefitter


Total Posts: 124
Joined: Oct 2007
 
Posted: 2015-05-01 23:32
fantastic. The only thing that could make it cooler is the ability of plugging in your own path finding algo.-

akimon


Total Posts: 566
Joined: Dec 2004
 
Posted: 2015-05-02 04:33
Thanks for the comments!

I added the ability to have two agents mate and pass on their genes to two offsprings (after crossover and mutations), after they survive for a certain amount of time. I think users would want to see this real time evolution effect.

You can also draw planks on the screen now with your mouse, and see how these agents react to newly placed planks set out to kill them.

The agents are controlled with only 10 neurons via input sensors surrounding its body. Of the 10 neurons, 2 controls movement. One controls whether it steers left or right (it cannot simply move straight, but must learn to do that), and another sensor controls whether it moves or stays still.


AndyM


Total Posts: 2321
Joined: Mar 2004
 
Posted: 2015-05-04 13:44
Fun stuff.

Funny how, when one doesn't have any agency, you cheer for the little guys, but as soon as you have the power to draw planks, you just want to decimate the little bastards.

I used to be disgusted; now I try to be amused...

NeroTulip


Total Posts: 997
Joined: May 2004
 
Posted: 2015-05-04 16:36
Are you thinking about applications to systematic trading or just playing around?

I guess we've all heard the stories about how these methods are useless for trading as they are too powerful at fitting the past and end up having no predictive value... On the other hand, part of me thinks this should work, and people who have tried this might have used the wrong training method, fitness criterion, or something like that.

If you put things in perspective on the Artificial Intelligence spectrum, most quant funds use things that are akin to "expert systems" right from the 80's. There's got to be something better. We're not talking about creating an Artificial General Intelligence here, being a good trader is a fairly narrow skill, more complex than chess / poker, but not that far away. How come we are so bad at it?

Inflatable trader

AndyM


Total Posts: 2321
Joined: Mar 2004
 
Posted: 2015-05-04 19:40
>being a good trader is a fairly narrow skill, more complex than chess / poker, but not that far away>

There's no leverage in chess :-)

Anyway, I'll take the other side of that proposition. Trading is mostly tacit knowledge, so I'm not surprised computers have made no headway. Even taking the example of chess, computers are strong simply due to brute force, not due to any real AI.

I used to be disgusted; now I try to be amused...

sharpend


Total Posts: 279
Joined: Aug 2007
 
Posted: 2015-05-06 17:44
Really nice. Thanks for sharing.

akimon


Total Posts: 566
Joined: Dec 2004
 
Posted: 2015-05-08 12:55
In the current state of machine learning algorithms, I think it is safe to say that AIs won't be able to become a Big Swinging Dick trader any time soon. More likely to replace an army of lower level execution trading grunts, or automate market making in a product with a deep order book. That is already being done, but with much simpler algorithms compared to neural nets because they work well (Most bank employees don't need to be smart anyways to do their job well, so why should the algorithms need to be that good?).

Neural networks are extremely flexible functions, and recurrent neural networks can even perform any computation a computer can perform as they are Turing Machines. However, training them to do anything useful requires a large amount of data in order to train the neural net to be able to generalise concepts and not overfit the data. For example, in this planks simulation, theoretically an infinite amount of data and scenarios can be generated so that the recurrent neural network can develop a strategy to survive. In the end, we are training the network to be good at performing something that is essentially deterministic, by giving it loads and loads of deterministic, non-random data.

Check out another neural net demo I made on swaption.net, if you haven't seen. The recurrent neural network has learned how to play a simple game with fixed rules and deterministic dynamics (in order to survive).

In my opinion, it would be difficult to train a neural network to predict future prices where the process is almost random anyways. Price history will look very random from the observer who does not have any other information or context other than pure historical price information.

From my experience, good traders are successful mainly from constantly trying to guess what other market participants will do in whatever scenario happens next, and position ahead of them, rather than, analysing a historical set of pricing data to predict future prices. Good traders also try to influence other market participants by trying to get others to trade something, which requires human interaction. Good traders are also great at assessing the impact of their actions on other market participants. So good machine trading systems will need to be more of a poker player, rather than a pattern recogniser. We are definitely not there yet, but perhaps we are very slowly getting there given the recent work done on no limit holdem, reinforcement deep q learning, and neural turing machines.


frstwrldprblm


Total Posts: 271
Joined: Apr 2008
 
Posted: 2015-05-08 18:55
Great post thanks!

NeroTulip


Total Posts: 997
Joined: May 2004
 
Posted: 2015-05-11 16:35
Actually, we're getting closer:

Poker bot

Inflatable trader
Previous Thread :: Next Thread 
Page 1 of 1