//****************************************************************************//
//*************** Production Systems - September 9th, 2019 ******************//
//**************************************************************************//

- *cue video of New Caledonian crow cracking a nut*
    - Clearly, these crows have some knowledge of the world: they know what operators are available, they know what their goal is, but there's a type of knowledge these videos don't talk about: "Operator Selection," or "Control" knowledge
        - This is how we choose what the right action is to get us closer to our goal! How do we know which operator to use at the right time? Crows can do it; how can we get robots to do it?
    - In fact, some researchers even use "action selection" as one definition of intelligence! (Although, like many things in this field, this isn't universally agreed upon)
--------------------------------------------------------------------------------

- In MEANS-END ANALYSIS, we say that one important type of thinking is selecting the right operator to solve a task
    - This is pretty straightforward if we can measure how far away we are from the goal state
        - In our block problem from last week, this is fairly easy: our "score" could be how many blocks are in the goal state!
            - Examining all our possible actions and their resultant state, this would lead to us putting C in the right spot
        - However, this leads to some difficulties: what if 2 possible actions have the same score? Do we pick one at random? Do we try both and see what happens? Do we give the robot a preference for some types of actions over others?
        - How do machines come up with these sorts of heuristics, anyway? For now, we'll engineer them ourselves - but we will talk about how robots themselves might handle this in the future
    - Notice, too, that there's an amount of metacognition happening here: we're evaluating our choices and "thinking" about them, and if it doesn't work out, we can go back and correct our mistake

- So, our next topic is PRODUCTION SYSTEMS
    - Professor Goel is...impressed that none of us have seen the video lectures for this topic
        - "You have to be watching the videos on these topics; I assume when giving these lectures that you've already watched the videos on Udacity covering these topics. Yes, there is some repetition, but I believe that repetition is useful"
    - What are these?
        - Imagine you're going to see a movie for the first time; how do you understand the movie? How do artists like The Beatles write songs?
            - As it turns out, people very rarely do it all in once step; if you're writing a song, for instance, you might first come up with the melody, then create the backing track, then write the lyrics, then finally put them all together
    - So, production systems are where we take a complex problem and break it down into several easier problems, and then solve each of those!
        - One example of this is "chunking," which many researchers consider the biggest piece of the human learning puzzle
    - How do we do that? Let's see!

- Let's think about baseball; how does a pitcher decide what type of pitch to throw?
    - This is again an action selection task, but there doesn't seem to be a clear heuristic for which one is best, and there are many different choices
        - How do they make these decisions? Similarly, how might a tennis pro decide what type of hit to return?
    - For instance, suppose it's the top of the 7th inning with 2 outs, there's a player on 3rd and 2nd base, and the current batter has a .252 batting average and was struck out last time. If the current score is 3-2 (us winning), should I walk the batter or try to strike him out?
        - We might represent this knowledge as a frame, like so:

            inning: 7
            porition: top
            runners: [2,3]
            outs: 2
            batter: Prado
            avg: .252
            score: 3-2
            goal: escape inning

- To help us with this, we might use our 3-layer cognitive architecture we talked about before:
    - The REACTION layer works within milliseconds, and tells us a decent action to take NOW
        - All animals have this layer, as far as we know
    - The DELIBERATION layer works over several seconds to several minutes, and involves learning new tasks, reasoning about what to do, and storing/accessing information we have in memory
        - This may involve working with several different types of knolwedge (e.g. procedural vs semantic knowledge)
    - The METACOGNITION layer involves thinking about thinking, and improving our reasoning processes
- Let's make a very bold claim now: "Architecture + Content = Behavior"
    - What are we proposing now? We're proposing that all intelligences have an ARCHITECTURE (a structure of mind/thinking), and CONTENT (the knowledge and different types of knowledge we know about)
        - As humans, we likely all have the same type of cognitive architecture, but the content within our minds make us unique

- So, going back to our pitcher example, there's MANY choices and potential outcomes
    - If we choose to walk the better, we have to solve the same problem for the next batter
    - If we choose to strike him out, we have to choose what type of pitch to use, and what might happen if a given pitch is a strike or hit or ball, and so on...
        - It's like we're running an adverserial simulation of the game in our head! Here, this type of thinking is about being able to generate possibilities before we make a decision!
    - We might try to solve this with procedural knowledge, where we have a set of rules on what to do in a given situation, and a way of choosing if multiple rules apply
        - If Professor Ashok is driving down the highway, for instance, he might need to decide if he's going to break the speed limit - what if he isn't in a rush? What if he is? What if there's a policeman?

- We'll come back to this next class - make SURE to watch the Udacity videos on this! It's important!