This process continues until proposed theories accurately predict any given design.

An artificial intelligence technique incorporating a specialized neural network is used to study observations, analyze parameters, propose theories,

generate hypotheses, and predict results. The neural network simulates both the structure and function of the brain. It is trained on a set of

observations (simulated wall designs) to predict results of future experiments (simulated wall designs) by analyzing causal relationships and

interactions among parameters.

The neural network goes through a process of logical reasoning and thought to make an educated guess, or hypothesis, of the optimal wall design.

The neural network predicts the price and minimum factor of safety of this wall design. The program then calculates a cost and a factor of safety for this

wall design. If the neural network prediction is incorrect, the neural network re-analyzes parameter relationships to generate a new hypothesis for the

best wall design. This process repeats until the neural network prediction of the hypothesis is correct.

The neural network then goes through a sleep stage, where the synapses connections release energy from the neurons. The neural network is then presented with a new random wall design, which the refreshed neural network uses in an attempt to think of a better wall design from a different perspective. The process concludes after no better design is discovered after several thought and sleep cycles.

This process is far more efficient than alternative minimization search algorithms, such as conjugate gradient, simulated annealing, or Monte Carlo techniques. Conjugate gradient descends through valleys in the search space until reaching a minimum, requiring numerous function evaluations,

and must be repeated for numerous random locations in the search space to increase the probability of finding a good solution. Simulated annealing

and other Monte Carlo techniques also require numerous function evaluations, but do not need to be repeated. Simulated annealing was implemented

for soil nail wall design as a comparison. Simulated annealing was able to find a good solution half the time after 12,000 wall design evaluations,

requiring about 10 hours. The neural network was able to find a good solution every time after 120 wall design evaluations, in about 10 minutes.

For example, a recent wall design from an existing project at a total cost of $35.4 M with a minimum factor of safety of 1.35 was analyzed by SNAP,

varying 9 parameters. SNAP generated 127 wall designs out of a possible 672 billion to obtain a VE design at a total cost of $21.0 M with a minimum 

factor of safety of 1.35, in 34 minutes. This was a cost savings of $14.4 M, or 41%, at the same factor of safety. The time required for the neural network training, logical analysis, and sleep cycles was 15 minutes. The time required to exhaustively search all 672 billion possible designs would be 345,000 years.