MuZero and the autonomous driving game

[Bet TSLA]

DeepMind's paper on MuZero is out in Nature. Here's their blog post:
MuZero: Mastering Go, chess, shogi and Atari without rules

For many years, researchers have sought methods that can both learn a model that explains their environment, and can then use that model to plan the best course of action.

So does this portend a future without traffic laws? If we move to a world where the notion of rules of the road disappears, and all that's left is goal oriented behavior, then this paper shows that we already have some idea how to deal with that.

If we have autonomous transportation software that can figure out the local conventions and preferences from scratch, then it can happily pilot pedestrians, bicycles, cars, trucks, delivery vans, tuk-tuks, etc. Why would rules be needed?

 

[powertoold]

Seems like MuZero is a master at perfect information environments.

Self-driving in the real world is a totally different problem. Not only is it an imperfect information environment, the environment can be arbitrary (like a chair falls off a truck).

Some people like George Hotz thinks we'll eventually have an end-to-end machine learning solution for self-driving, but I have trouble seeing how that can happen (unless we've already achieved general intelligence). Then again, I'm not a machine learning specialist.

 

[jboy210]

Interesting question. Machine Learning models can be trained from experience and have an exploration component that tries new things that most of DeepMind's Reinforcement Learning models do. The issue is what happens when the rules or conventions change. Since ML models are experience-based, if they can train it with enough data it can adapt to changes in the same way we humans can.