Importing a trained artificial neural network to a particle swarm optimization function

I already have a trained ANN which I would like to incorporatie into a function which optimizes using PSO.

[Yusuf Shakeel]

In my ANN, I had 45 inputs in total and 1 output (efficiency). This ANN has been created. Now I want to use this trained ANN and PSO to solve this problem:

• 39 of the inputs are fixed. I have already initialized and normalized (mapminmax) these.
• 6 of the inputs will be chosen by PSO with the help of the trained ANN so that the efficiency (output in ANN) is maximized.
• Therefore the output of the PSO function shall be the 6 variable inputs (inputs in trained ANN).

I am unsure about how to call the trained ANN and how to input the 39 fixed and 6 random variable inputs to the PSO function so that it can return the (6) optimized variable inputs.

Answer 1

[Marcus Södergren]

I do not get what your question is. Would you mind to specify?

My first thought, though, was that you should implement a PSO where you initialize your particles in the neigbourhood of your already trained network, and with randomly chosen velocities. That way you can make use of the fact that the network is already trained.

[Yusuf Shakeel]

Thanks for responding.

In my ANN, I had 45 inputs in total and 1 output (efficiency). This ANN has been created. Now I want to use this trained ANN and PSO to solve this problem:

• 39 of the inputs are fixed. I have already initialized and normalized (mapminmax) these.
• 6 of the inputs will be chosen by PSO with the help of the trained ANN so that the efficiency (output in ANN) is maximized.
• Therefore the output of the PSO function shall be the 6 variable inputs (inputs in trained ANN).

I am unsure about how to call the trained ANN and how to input the 39 fixed and 6 random variable inputs to the PSO function so that it can return the (6) optimized variable inputs.

Hope I made it clear.

[Marcus Södergren]

Ok, I think I follow.

The PSO algorithm looks something like:

1. Initialize
2. Evaluate fitness for all particles
3. Uptade the best position values
4. Update particle velocities
5. Update particle positions
6. Repeat 2-5 until termination condition is met

So you will run the positions of the particles through the ANN when you calculate your fitness (a sutable fitness might be fitness=1/ANN-output).

Does this help you?

[Yusuf Shakeel]

Yes thank you.