11/06/2020
Discipline without Play is useless. It is so in Art; in Design and in biological evolution. And AI confirms it.
A few days ago, I asked a question on Linkedin that I expected no one to pick up :D. I read a post on being a result of the executive-attention network (which refers to the process of learning and achieving mastery) and the imagination network (which is a phase of exploration and play) and I thought that all systems I knew about, seem to benefit from randomness. For example, in social media, when the system only makes recommendations based on user's previous choices, it runs into the danger of overfitting (and this creates information silos where we all see what we like but we end up not understanding others!). So I was loudly wondering if there is a model to describe how much randomness we need in a system. Then a friend linked this article that explains the "steppingstone" principle. It turns out that in a simple navigation task, if we reward robots for novelty, we'll produced more frequently a solution to the problem compared to rewarding them for performance. Yet, there are caveats. It's an exciting read with so many implications.
And it reminded me of "Why bad ideas are a good idea" of Alan Dix. Good ideas can take you only to their local maxima (which might be good enough!). Bad ideas extent your solution space.
By ignoring their goals, evolutionary algorithms have solved longstanding challenges in artificial intelligence.