Summary

In this chapter, we developed a generic MLP module, using formalized interfaces:

• The perceptron simulation interfaces are left separate from the training interfaces so that there is minimal overhead for nonlearning neural networks.

• The batch training algorithms also require more memory, and are separated.

We designed an animat capable of gathering information from the environment to extract facts about its rocket shots. This enabled us to learn to predict the damage of each rocket. The animat can therefore select random targets, and use the perceptron to estimate the best:

• Gathering the information from the environment to monitor the rockets was the biggest challenge.

• To deal with the noise in the data, the examples were stored in a common file.

• Plotting graphs and histograms of the different features revealed that the distances are the most reliable inputs.

• The target selection is best learned offline by a batch algorithm.

Thanks to the perceptron, the AI developer spends less time solving a relatively complex problem. The perceptron exploits learning technology to improve the aiming capabilities beyond hand-crafted solutions. The MPL has the advantage of being easily adapted to the skills of different animats, too.

The next chapter presents an important concept: understanding the problem. This is an essential skill for AI developers because understanding the problem always drives AI system design.