What am I missing here, because I'm positive that I am missing something?
[0]. https://youtu.be/PUv66718DII?t=29m20s [1]. https://youtu.be/ZfytHvgHybA
The set of facts is worthless for anything of complexity. It does not really generate the rules itself. (They are directly derived from the facts.)
What they did is only a small improvement over a typical expert system or CNN for a very limited case.
Choice quote: "Notably each fact can be linked back to the characteristics of a sprite that it arose from." Wrong. When you pick up a flower your sprite changes, but how does it know you can suddenly shoot bullets? Etc. And for more complex games a lot of data requires exploration well past the GUI. An action might change acceleration (suddenly nonlinear ice physics with momentum), or direction handling, or you can start flying, or many other things. What if the thing moves in a circle? What if there is just some probability that something results?
The approach will fail at modelling as soon as Mario level 1-4. (The one with rotating fireballs.) Or produce an insane representation of the engine. Note how it even cannot model the dampened triangle wave motion of the fireballs in the example - assumes they're a sparse line.
The paper presents no way to reduce these huge number of "if-then rules" to something actually useful either.
Since this doesn't even attempt to explore the state space, it also requires a huge database.
Calling this "recreate game engine" is akin to saying that since we have an algorithm that can solve checkers, it will solve poker, go and also whodunit. And can play Jeopardy too.
I even suspect it's not useful as a preprocessor to something that can actually play a game, as it will break later cases.
CNNs have done such impressive things that "outperforms convolutional neural nets" sounds like an achievement, but CNNs have never been the pinnacle of accuracy - their key advantage is flexibility. Feature learning costs some reliability, but gives a huge advantage in saving human time and effort.
This appears to be exactly the opposite approach, an AI system that gains its accuracy by working from heavily pre-defined rulesets. Feature engineering is fine in a stable, well-understood domain, but it reduces the impressiveness of the 'AI' result. And more worryingly, it cripples the flexibility of the agent in a open domain like "video games".
Hand-authoring a set of functions required to derive the model means embedding a huge portion of the game engine in the engine-learning framework - what's left to learn is basically just parameter values. Mario without powerups is a game entirely defined by 2D movement, collisions, animation, and a tracking camera. That's the same feature list that had to be hand-defined for the engine.
I don't mean to attack the authors. This is still an interesting result, and they do acknowledge this in P2 of 'Limitations'. (Albeit with some lofty claims about eventually understanding real video - are they planning to encode physics as their ruleset?) But the article really oversells the capacity of a system that was spoon-fed the essentials of what it had to learn.
So what you're saying is they came up with some new, albeit small, way to do it?
Yes, the technique it uses only works for a certain space of possible games. That means there is an obvious path to increasing the size of that space.
In addition, this is wrong to having been said to be new, such attempts have been made before and even stronger in results and generality. For example this (relatively dumb) approach from 2013 generalized kinda well, much better than I've seen a silly even deep network generalize: http://www.cs.cmu.edu/~tom7/mario/
So yes, they are overselling it a lot. I am 100% not impressed by this paper as it lacks critical detail. That it can parse stuff from 2D frames is not interesting, it is basic motion analysis which can be done even by a supremely stupid algorithm, not even a CNN.
I mean, Google best AI can play 15 rooms of a simple game...
(The algorithm as described will require a huge database for a game that is even slightly more complicated than Infinite Mario. And we don't even have the sources to try that.)
Even the object motion tracker part will choke in 3D environment. (It is a greedy matcher as they described it.)
Speaking of impressed, Google DeepMind paper is way more actually feasible to improve upon and rich in detail: https://arxiv.org/pdf/1606.01868v1.pdf Compare the two papers in straight quality. I understand why you'd publish any worthless junk in the current academic culture and do not agree we should actually do it.
Section 3.1 of the paper outlines a list of 'hand-authored' functions the agent used to derive events from images. They include animation, sprite-entity relationships, motion, collision, and camera movement. Which is to say, every component of Super Mario level 1-1.
That doesn't mean the paper is uninteresting, or useless. Defining facts based on those possible rules is still an intriguing result. I'm having real trouble working out from the paper how well their agent understood conditional changes like size and fire flowers - if it accurately recreated those rules, then I am impressed.
But "modeled without accessing the code" is a dubious claim about an agent that started with a list of the core rules included in its code. The Engine Learning section (3.2) mentions that automatic derivation of possible facts is a key area for future work. That is to say "this would be flexible if it did feature learning instead of needing feature engineering". Unfortunately, that's the problem in agent design, and the value of CNNs isn't unbeatable performance but the capacity for flexible feature learning. The press release here elides the issue of feature learning entirely when comparing performance.
The idea of producing a rule-based system from deep-learning, while not exactly a breakthrough, is an interesting direction to take.
It is research. It is not designed to solve real-world problem but to give ideas to engineers. And really, I can see several simple systems that could be programmed by simple rules and learned from input/outputs.