I would like to add more example programs to the GAUL distribution. But, I don't want to waste my time writing dumb examples when I could be wasting my time solving a useful (or at least semi-useful) problem instead.
Any ideas for example programs out there?
I'm tinkering with the idea of writing a neural net that's trained using a GA instead of plain back-propagation. In the scientific literature, one or the other tends to be used, but I see no reason why a hybrid method wouldn't work. My plan is to use a Lamarkian GA, where the adaptation is back-propagation. I'll knock something together when I have time. Comments?
Stewart.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
If anybody is interested, a neural network evolution example following the scheme suggested is now included in the distribution.
see: examples/nnevolve.c from the source code package.
In this example, a fixed topology neural network is trained using a combination of back-propagation with momentum and a GA. The GA portion modifies the synaptic weights and also the learning parameters. The optimal set of learning parameters is usually problem specific, but this issue is side-stepped by the GA scheme.
I admit that this example is a bit silly, because the standard back-propagation algorithm can quickly do a good job of training the network with the provided data.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Isn't one of the advantages of GA trained NNs to sidestep the local minima issue?
I am looking for a GA library for a trading application. I already wrote a simplistic GA, actually it was a crude GP generating evaluation formulae in the style of:
(Close[1]>Open[2]) AND (...) OR NOT (...) ...
Rather than integrating some old proof of concept code into my current development I am considering using an outside library so I can focus on gene definition.
If you want to have some idea of what I am trying to accomplish you may check:
> Isn't one of the advantages of GA trained NNs to sidestep the local minima issue?
True. Although, I now think that the are better (in terms of training speed) approaches than this GA idea for jumping out of local minima. What I've tended to do recently is train the network, randomly perturb many of the weights, and repeat until no improvement is seen.
I've fairly busy at present, but what language would you like the wrapper for?
HelGA? That's some protein modelling software which I wrote during my PhD studies. It made heavy use of GAs. The GA code became GAUL...
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
> I am looking for a plain Wintel DLL, C or Pascal calling conventions.
I see no reason why GAUL couldn't be compiled into a windows DLL without too much hassle. However, I have no easy access to any PCs running windows. Do I hear any volunteers?
> Would GAUL be a good match for a binary Tree GP implementation?
I suppose so. I have never tried this though.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
The newspapers here started publishing a number puzzle called sudoku (see www.sudoku.com). My feeling is, the problem could be solved in GA (like 8 queens problem). Would you like to look at this as an example?
Ahmet Sekercioglu
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
I would like to add more example programs to the GAUL distribution. But, I don't want to waste my time writing dumb examples when I could be wasting my time solving a useful (or at least semi-useful) problem instead.
Any ideas for example programs out there?
I'm tinkering with the idea of writing a neural net that's trained using a GA instead of plain back-propagation. In the scientific literature, one or the other tends to be used, but I see no reason why a hybrid method wouldn't work. My plan is to use a Lamarkian GA, where the adaptation is back-propagation. I'll knock something together when I have time. Comments?
Stewart.
If anybody is interested, a neural network evolution example following the scheme suggested is now included in the distribution.
see: examples/nnevolve.c from the source code package.
In this example, a fixed topology neural network is trained using a combination of back-propagation with momentum and a GA. The GA portion modifies the synaptic weights and also the learning parameters. The optimal set of learning parameters is usually problem specific, but this issue is side-stepped by the GA scheme.
I admit that this example is a bit silly, because the standard back-propagation algorithm can quickly do a good job of training the network with the provided data.
Isn't one of the advantages of GA trained NNs to sidestep the local minima issue?
I am looking for a GA library for a trading application. I already wrote a simplistic GA, actually it was a crude GP generating evaluation formulae in the style of:
(Close[1]>Open[2]) AND (...) OR NOT (...) ...
Rather than integrating some old proof of concept code into my current development I am considering using an outside library so I can focus on gene definition.
If you want to have some idea of what I am trying to accomplish you may check:
http://mapage.noos.fr/arbitragex/current_projects.htm
I am not a C programmer. Any time table for the wrapper?
PS: What is HelGA?
Hi Terry Domps
The link provided by you is not working.
http://mapage.noos.fr/arbitragex/current_projects.htm
Anyway I am interested and working in the similler field that you are developing and willing to use GAUL for this purpose.
Do you have any working model which we can go forward for developing it further.
I have knowledge of C and C++ and can use MPI as well.
Regards,
Pawan
> Isn't one of the advantages of GA trained NNs to sidestep the local minima issue?
True. Although, I now think that the are better (in terms of training speed) approaches than this GA idea for jumping out of local minima. What I've tended to do recently is train the network, randomly perturb many of the weights, and repeat until no improvement is seen.
I've fairly busy at present, but what language would you like the wrapper for?
HelGA? That's some protein modelling software which I wrote during my PhD studies. It made heavy use of GAs. The GA code became GAUL...
I am looking for a plain Wintel DLL, C or Pascal calling conventions.
Would GAUL be a good match for a binary Tree GP implementation?
> I am looking for a plain Wintel DLL, C or Pascal calling conventions.
I see no reason why GAUL couldn't be compiled into a windows DLL without too much hassle. However, I have no easy access to any PCs running windows. Do I hear any volunteers?
> Would GAUL be a good match for a binary Tree GP implementation?
I suppose so. I have never tried this though.
I have been told that if you install cygwin ( http://www.cygwin.com/ ) and use that to compile GAUL a set of DLLs are automagically generated!
Indeed. It allows one to use gcc on the PC. I have not read anything yet that suggests how/if this set up will compile C code into a wintel DLL.
I somewhat doubt it will but I will take the risk to get flammed and ask on the cygwin mailing list ...
Hi Stewart,
The newspapers here started publishing a number puzzle called sudoku (see www.sudoku.com). My feeling is, the problem could be solved in GA (like 8 queens problem). Would you like to look at this as an example?
Ahmet Sekercioglu