Optimizing a Multi-Model Reward System

In a system that utilizes an ensemble of reward models to generate a final reward signal, what is the key analytical advantage of employing a fusion network for combination, as opposed to a simpler method like calculating the mean of the individual model outputs?

You are tasked with implementing a system that combines outputs from an ensemble of different reward models to produce a single, refined reward signal using a specialized neural network. Arrange the following steps in the correct chronological order to design, train, and deploy this network.