Thursday, 25 October 2018

perfect play or macky, micky, mucky, mocky

As an experiment to explore how norms and ethics might be introduced to artificial intelligence in a broad and general fashion, researchers, as Slashdot reports, have trained implicitly one such programme to play Pac-Man and win without gobbling up the ghosts.
The training was a balance between the programme’s drive for optimisation tempered with lessons from human players that avoided the ghosts, even at their most vulnerable, and eventually netted more ethically informed play. It’s not quite the level of trust that I would want in a moment of pursuit but I suppose it does illustrate the potential to build in moral false-safe measures.