next up previous
Next: Evolution as Mixture of Up: Discussion Previous: Discussion

Adapting to the Real Problem

We use a state-of-the art evolutionary algorithm, continuously being trained by self-play, as our novelty engine; but the players generated have a broad RS distribution: some of them are very good against humans and others are very bad. In all, even with feedback from the human results (by using champions as trainers, or later, by allowing champions to reproduce) the average RS of the robots produced is below the performance of Tron as a whole (section 3.7).

A Tron playing system relying solely on self-play would not be able to improve beyond this rate. It was the power of selection coming from interaction with a live human population that drove the system up to an expert level.

Pablo Funes