Hands-On Neuroevolution with Python

The novelty metric

The novelty metric is a measure of how different the current solution is from the already known ones. It is used to calculate the novelty score when estimating the distance from the current point in the behavioral space to its k-nearest neighbors.

In our experiment, the novelty metric measuring the difference in the behavior of the two agents is determined by the item-wise distance between the two trajectory vectors (one vector per agent). The trajectory vector contains the coordinates of the positions that were visited by the maze navigator agent during a simulation. The following formula gives the definition of the metric:

Note is the size of the trajectory vector, and and are the values at position of the compared ...