I am a beginner in using joone APIs for building neural network. I have a problem and I am planning to use neural network.
My data is something like below. Each input is separated by semicolon for joone APIs file.
For simplicity I gave one few inputs. Puts can be as high as 40.
Input: 1;-20;2;-40;3;-30;4;-50;
output: 1 (or binary represented as 0001)
Input: 1;-20;2;-80;4;-50;5;-100
output: 3 (or binary represented as 0011)
My problem is for a given input, I must find a close match to the output. Input could differ little bit and network is smart enough to pick best output.
I started from XOR problem and it worked. I modified the program for this problem. I get high error rate and not useful. I am using sigmoid transfer, full synapse, 1 input + 2 hidden layers + 1 output layer, number of neurons in hidden layer is same as input, momentum is 0.3, learning rate 0.8, number of output neuron 4 for binary representation of output. I tried using many hidden layers like 10 or 20 and didnot help and now at 4. I tried epochs from 1000 to 10000. For each iterartion the error rate is reducing but only at 5 decimal digit or more. So very slow learning process.
Can someone experienced guide me on what I can try to improve. Error rate now is 0.727.
Any suggestions on changing the network parameters or how to approach to solve this is appreciated.
Thanks,
Rajesh
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi there, did have tried to change number of neurons in your hidden layer?
test a RBF network with just one hidden layer and gauß-function as activation function. maybe it will help you.
best regards
puyan
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I am a beginner in using joone APIs for building neural network. I have a problem and I am planning to use neural network.
My data is something like below. Each input is separated by semicolon for joone APIs file.
For simplicity I gave one few inputs. Puts can be as high as 40.
Input: 1;-20;2;-40;3;-30;4;-50;
output: 1 (or binary represented as 0001)
Input: 1;-20;2;-80;4;-50;5;-100
output: 3 (or binary represented as 0011)
My problem is for a given input, I must find a close match to the output. Input could differ little bit and network is smart enough to pick best output.
I started from XOR problem and it worked. I modified the program for this problem. I get high error rate and not useful. I am using sigmoid transfer, full synapse, 1 input + 2 hidden layers + 1 output layer, number of neurons in hidden layer is same as input, momentum is 0.3, learning rate 0.8, number of output neuron 4 for binary representation of output. I tried using many hidden layers like 10 or 20 and didnot help and now at 4. I tried epochs from 1000 to 10000. For each iterartion the error rate is reducing but only at 5 decimal digit or more. So very slow learning process.
Can someone experienced guide me on what I can try to improve. Error rate now is 0.727.
Any suggestions on changing the network parameters or how to approach to solve this is appreciated.
Thanks,
Rajesh
Hi there, did have tried to change number of neurons in your hidden layer?
test a RBF network with just one hidden layer and gauß-function as activation function. maybe it will help you.
best regards
puyan