Once again, we initialise Q as zero on every link and allow the system to reach equilibrium.
In addition, observe that we require the knowledge of the number of layers N. In our approach we initialise this value using a stereo match algorithm [20].
For case E we initialise (B^{prime y}=B^{y}/2) and case F is initialised similarly as (B^{prime x} = B^{x}/2), (B^{prime y}=B^{y}/3).
In the algorithm, we can implement this by defining an n-dimensional vector S. We initialise this vector on S j = 0 and for each base station that is switched on, we put the value on 1.
Here we initialise the parental and offspring 6-node network layers with weak all-to-all connections, i.e. within each layer, neurons are fully connected with random weights in the range 0 and 1 mV in magnitude.
We initialise the GA population with randomly generated valid structures using the procedure shown in Algorithm 7. As noted before, our GA uses a macro-mutation operator that tries to take the search towards forming a hydrophobic core.
For each probabilistic local walk starting at a seed s, we initialise the current vertex to be the seed s with an initial energy of α in lines 3 and 4. In lines 5-14, the algorithm walks from vertex to vertex until the energy falls below 0. At each non-seed vertex that it visits, it increments visitCount s, v).
To evaluate Φ ( V 0 ; μ 0 ) we first lift V 0 to produce v 0, where L ( V 0 ) = v 0. We then initialise the fine-scale model (70) with the initial condition v 0 (and μ = μ 0 ) and simulate (70) for a short amount of time, of duration say Δ, resulting in the state v.
Starting with the DNN models trained using foreign data, the output layer is replaced by a new layer in which we randomly initialise the W L, which is the matrix of connection weights between the layer L−1 and the output layer L. The network is then retrained using the MP-FR data, which most closely matches the evaluation set.
