-1

From everything that I have seen about neural networks and genetic algorithms, I have noticed a few things :

NNs are very good at converging to a solution after a certain number of iterations. GAs are good at finding a solution to a problem after a certain number of generations. However there is one time-complexity obstacle at play here, and that is the actual construction of the neural network and genetic algorithm itself. This is where the actual skill and understanding comes in : the consideration of nodes, weights, activation functions etc. For genetic algorithms it's the fitness function, error values etc. This is all determined by the problem domain itself.

My proposal is to find a generalized algorithm that can take the problem statement such as "create a program that simulates a netball game and finds optimal strategies for effective play", and CREATE THE NEURAL NETWORK OR GENETIC ALGORITHM ITSELF using a combination of databases, statistics, classification systems, logic, decision theory, mathematics.

One possible approach to at least partially solve the problem is having a database of problem domains, and an existing neural network and ga for that problem.

The database could have the following attributes :

Problem statement : VARCHAR, 
Problem domain    : VARCHAR, 
numLayers         : INT,
NNTree            : TREE, 
numNodes          : INT, 
activationFunct   : LIST

As the user specifies the program statement, the program must break it up into its elements. For eg , "Netball simulator that learn the effective strategy of the game",

is broken up into Netball [ Rules of game are known ], simulator [ implies 2d or 3d graphics, predefined objects for graphics ], effective [ interpreted as optimal, which influences that activation function used], strategy [ interpreted as emergent behaviour of objects ], game [ interpreted as goal oriented action list ]

What data structures or algorithms are needed for this ?

AJ.
  • 4,526
  • 5
  • 29
  • 41
York Cleave
  • 1
  • Possible duplicate of [Clarification on a Neural Net that plays Snake](http://stackoverflow.com/questions/42099814/clarification-on-a-neural-net-that-plays-snake) – devinbost Feb 15 '17 at 20:34

1 Answers1

2

If someone could solve this problem, he would actually be creating intelligence out of nothing. There is no (and in my humble opinion will never be) a way of automatically discover a way to select a solving problem method. This would presuppose that the problem's solution would already be known. In such case, there is no point to discover the solving problem method...

Hope I helped!

Pantelis Natsiavas
  • 5,293
  • 5
  • 21
  • 36
  • The process that a human uses to decide what the architecture of the neural network is,that is the process that I am trying to simulate. It is actually closely related to decision theory, which apparently can be extracted from existing neural netorks. Although I haven't found software yet that can do that, unless you know of any ? The algorithm to create the neural network itself, is based on probability, it's not an exact science. The point is if you have a database of neural nets, then you can borrow from those nets according to their similarity, to create the new neural network. – York Cleave Nov 22 '13 at 07:36
  • 1
    Pantelis is right here. IMHO you are simply aiming to high. Much smaller problems are too hard to be solved be teams of great researchers. One thing, that should show you what are you trying to "attack" is so called "meta learning", just read some papers from this field and you will probably understand how hard is this problem. – lejlot Nov 22 '13 at 18:41
  • York, if you have seen something similar like the approach you propose, please give a reference. I would really like to see how a decision process of what neural network to build or what fitness function to use is modelled. I am sure that we could build a data model, if the conceptual model is clear. Right? – Pantelis Natsiavas Nov 22 '13 at 19:16