[globuous]

Agreed. I much prefer we would call it statistical intelligence.

Although artificial intelligence is actually spot on. We just understand the wrong side of the ambiguity. Its not really intelligence that we have reproduced artificially - since it isn’t intelligence - but a fake intelligence, the artificial kind. We’ve created the artifice of intelligence, through statistics, but not intelligence.

People knew long before newton that an apple would drop to the ground when released. Statistical experience has allowed us to have knowledge of this very early on. But it took newton to explain what was going on, so that instead of predicting through experience, he could predict by reason and logic. Thus saving him many lives of accumulating experience to make his next prediction ever more precise.

"Statistical intelligence" allow us to do a bunch of neat things though. Many problems are best approached statistically (because noise, lack of formal understanding etc), and these some of these methods achieve impressive results in a wide range of situations.

[nathanaldensr]

The use of the word "intelligence" overall is problematic. To most people, "intelligence" is human inductive reasoning. We see "intelligent" creatures in mass media and books--creatures that act just like humans except aren't biological. We think of Commander Data from Star Trek. Proponents of AI know most people interpret the term that way and gladly use the term as a way of implying the same magic we see in media.

[streetcat1]

So advances in RL (Deepmind), are not merely statistical intelligence, those are true advancement in AI (not only ML). I.e. those a machine can train on their own data.

[globuous]

True but i’ll argue a bit. They statistically maximize reward. As far as i’m aware, the engineer is still designing the reward function. She’s also designing the statistical method to converge to the optimal solution (as quickly as possible).

So a RL chess algorithm tells your statistically a move (action) from a state S to a new state S’ such that you are expected to maximize your reward. Whereas a chessmaster (probably) designs his next sequence of moves based on logic (my opponent will respond in such a way because etc). This is different from « statistically, this move right now has the best odds of leading to a win » a la monte carlo. Now what is surprising, is that statistical algos are better than our best logicians at this particular task. But its the action at a given state is still statistically designed.

Finally, you need your data mining to be representative of the underlying distribution you are trying to model. So you need your simulator to be the most real whereas they are in fact approximations in most useful cases (landing a plane for instance).

So for instance if you want an algo to design the flight path of a rocket landing on an asteriod, you could recreate a simulator modeling spacetime from observations and model its dynamics from eintein’s equations, but then what’s the RL for, why not just use an off the shelf optimization algorithm like we have for decades? [1]

The bellman equation and DQNs are nice and all, but they’re still statistical algorithms, producing - in my mind - statistical intelligence about a particular system. An RL agent will not tell you WHY such an action was taken, but it’ll tell you that statistically, it is the action to take.

Very neat results in RL however.

[1] i worked on a RL based agent to control trafic lights, and it wasnt clear whether our solution was better than a classical optimization one. Actually, classical optimization (minimizing an analytical model of the system) seemed to scale much better to larger meshes.