Slashdot Mirror

← Back to Stories (view on slashdot.org)

Ask Slashdot: Best Way To Learn About Game Theory and AI?

Posted by Soulskill on from the watch-scifi-movies-from-the-eighties dept.

xmojox writes "I would like to learn more about Artificial Intelligence and Game Theory. I know these are both large areas of study; however, my main interest is in how these affect decisions in the world. This would include politicians, business people, and general society. I'm not looking for a career or anything; this is just a personal interest of mine. Where are good places to start in these areas for somebody new to them? I'm aware of the Stanford on-line classes, but those don't work with my current schedule."

32 of 152 comments (clear)

  1. Russell and Norvig by Anonymous Coward · · Score: 3, Informative

    Grab a copy of Russell and Norvig. It's a nice survey, and a fairly easy read.

    1. Re:Russell and Norvig by crow_t_robot · · Score: 2

      Seconding this. Norvig is also teaching a free class online next month from this text through Stanford.

    2. Re:Russell and Norvig by somersault · · Score: 2

      Good job on reading the entire summary!

      --
      which is totally what she said
    3. Re:Russell and Norvig by OeLeWaPpErKe · · Score: 2

      No the correct answer is : "You want to learn the basics for no other use than misrepresenting it in political discussions ? Please don't."

  2. statistical decision theory by ThorGod · · Score: 4, Interesting

    -bone up on your probability (continuous/discrete distributions, transformations, etc)
    -grab a book on statistical decision theory like Parmigiani and Inoue or Berger (85).
    -read Von Neumann/Morgenstern

    --
    PS: I don't reply to ACs.
    1. Re:statistical decision theory by scottgfx · · Score: 2

      Many years ago I read the book "Prisoner's Dilemma".

      Interesting book with a bit of Game Theory and biography of Von Neumann.

      Thought it interesting to note that my father's and Von Neumann's lives ever-so-slightly intersected at Operation Crossroads.

      --
      It's mandatory to wash your hands before returning to the land of Dairy Queen.
    2. Re:statistical decision theory by Slugster · · Score: 3, Interesting

      -read Von Neumann/Morgenstern

      I have the Von Neumann/Morganstern book. It is very heavy reading, Rain-Man level stuff. Unless you're rich or its really cheap, it's a good idea to thumb through a copy before buying.

      On the other [fuffy] end of the spectrum is Prisoner's Dilemma by William Poundstone. A 1-2 hour read suitable for teens, with no difficult math and a lot of real-world examples.

    3. Re:statistical decision theory by sammyo · · Score: 2

      The first chapters of Theory of Games and Economic Behavior is quite accesable, at least it was when I got my copy in high school. Do to be able to grasp the idea of a matrix.

  3. Less Wrong by Bob9113 · · Score: 2, Interesting

    I haven't had much time to dig in yet, but I hear good things about Less Wrong from some friends who are into game theory, ai, and sociology.

    Here's their front page blurb:

    Thinking and deciding are central to our daily lives. The Less Wrong community aims to gain expertise in how human brains think and decide, so that we can do so more successfully. We use the latest insights from cognitive science, social psychology, probability theory, and decision theory to improve our understanding of how the world works and what we can do to achieve our goals.

    --
    Stop-Prism.org: Opt Out of Surveillance
    1. Re:Less Wrong by GrimmParoD · · Score: 2, Funny

      You should be careful with communities assembled around prohibited subject matter. Game theory is one thing - Singularity class AI would be so disruptive that it may be assumed unregulated advances in the field could get you 'cleaned up' in pretty short order. People peddling entre into communities gathered around such subjects should be considered suspect unless they openly tell you that any significant contribution will likely come with total loss of your freedoms.

      This fact gets conveniently left out of literature from Singularitarians in particular, but should be logically apparent to anyone actually deserving of being included in such an effort. If you must work on AI, either work alone and air-gapped, or alone and on a machine from which you periodically notify the NSA of your intentions to overrun the world with sentient kill-bots. Both of those options are better than walking into what should be an obvious death-trap.

      The existence of security based prohibitions may suck, but it doesn't increase freedom to associate with individuals who are so obviously positioning to catch indies in a highly regulated field.

      My advice, forget 'game-theory'/AI terminology and work on non-verbal thought processes via extended meditation. If you must use language, develop your own compaction routines with cipher keys bundled for obviousness - don't resort to natural language when attempting to make a leap across semantic boundaries. Trying to separate the expansion of logical processes from the compaction of logical processes is largely useless, and the security bump from obvious behaviors will pay off if you get popped for making progress.

      Better to go get laid and have kids if you want to study emergent systems. Just MHO. Now where did I misplace my breeding stock? Hrm... Not here in Mom's basement.. *wanders off*

    2. Re:Less Wrong by Hatta · · Score: 2

      Right at the top of the page they have a line endorsing human rationality. Right next to that is a link to a singularity summit. There is no hint of irony in the juxtaposition. That's a big red flag right there.

      --
      Give me Classic Slashdot or give me death!
  4. How do they not work?? by SuperKendall · · Score: 3, Insightful

    I'm aware of the Stanford on-line classes, but those don't work with my current schedule

    Why? You can just watch the videos instead of doing the homework, or watch them sometime later and do the homework then.

    But if you really had any interest you would be shifting around everything else, including sleep, to take fullest advantage of these classes in real time.

    --
    "There is more worth loving than we have strength to love." - Brian Jay Stanley
    1. Re:How do they not work?? by fartrader · · Score: 2

      Its a good thing that everyone seems to be clairvoyant and understand how the OP *can* indeed fit it in to his current schedule when he says he can't. Maybe he has a 100 hour a week job, a demanding girlfriend, 200 kids or all three. ...and "interest" doesn't mean you give up sleep. Perhaps if he was interested in two things he should give up eating as well.

      Reasonable question asked - with reasonable parameters - unreasonable dissection.

    2. Re:How do they not work?? by elsurexiste · · Score: 4, Insightful

      Why? You can just watch the videos instead of doing the homework, or watch them sometime later and do the homework then.

      Maybe xmojox isn't around a PC for hours, because of his/her job, and spend a few more hours commuting. Maybe he/she has no tablet that's easily carried. We are not to judge. If we don't believe what's telling us, we may as well think this question is just an attempt at trollling and stop wasting our times.

      But if you really had any interest you would be shifting around everything else, including sleep, to take fullest advantage of these classes in real time.

      Worst. Advice. Ever. If you don't sleep at least six hours, you'll notice. It isn't sustainable.

      On topic: Check Wikipedia's page for Game Theory and go to the citations. There you'll find a few books and other resources you can read.

      --
      I rarely respond to comments. Also, don't ask for clarifications: a brain and Google are faster, believe me!
    3. Re:How do they not work?? by fartrader · · Score: 2

      ah that's because unlike the other posters I am in fact clairvoyant :)

  5. Game Theory - The Great Courses by slasher999 · · Score: 2

    I purchased a course from "The Great Courses" on DVD last year (thegreatcourses.com), the topic of which was Game Theory. I've enjoyed the first half of the course, but haven't completed it. Unfortunately whenever I get time to go back to it, it has been long enough that I tend to start back at the beginning and watch the entire course over.

  6. Russell & Norvig by Black+Parrot · · Score: 4, Informative

    Read Artificial Intelligence: A Modern Approach, 3rd edition. It's supposedly the most-used AI textbook in the world.

    It's weak on the biologically inspired methods (genetic algorithms, neural networks, fuzzy logic), but very solid in "Good Old Fashioned AI" (GOFAI) and some of the decision-making procedures from other fields such as economics.

    If you don't have a background in CS, you'll need to work through a book on discrete math first.

    --
    Sheesh, evil *and* a jerk. -- Jade
  7. The best way to learn is to do it by Unoriginal_Nickname · · Score: 3, Interesting

    The best way to learn is to do it. Choose a "game" and try to solve it with some different approaches. I say "game" with quotes because the game you pick should definitely not be a game which a normal adult would choose to play, but something very young children would play, or a heavily simplified variant of a full game. Something like Tic-Tac-Toe or RPS.

    RPS seems trivial, but it's actually a very interesting game to study. It's an easy-to-understand example of how a Nash equilibrium strategy doesn't always produce an optimal outcome. The equilibrium strategy is to choose between the three moves at random, but you can't naively use the strategy because it offers no way of taking advantage of weak opponents, such as an opponent that favors a particular move or a pattern of moves. Computer RPS tournaments will always include a variety of bots that are predictably weak in various ways, to separate out the good bots that are capable of using these weaknesses.

    Another simple game you could experiment with is Leduc Poker. Leduc Poker is another matrix game, and it's simple enough that you can easily compute the Nash equilibrium (which, remember, is not necessarily optimal, but it's a good starting point) or iterate over the entire game tree. You could also use a similar subset of poker to experiment with more advanced techniques - e.g. minimax and alphabeta pruning, or maybe Monte Carlo Tree Search (I can't guarantee that MCTS would work for poker, I'm not sure it's ever been done, but it might be interesting to try.)

    1. Re:The best way to learn is to do it by Rockoon · · Score: 2

      In all likelihood, the OP doesnt actually want to learn AI but rather the related subject of Machine Learning.

      The problem with tackling AI is in fact everything you were talking about. Pick a simple game, write a player for it, etc, where the answer is nearly always some form of tree search leveraging hard-coded knowledge (chess-like) or simple Bayesian derivations using hard-coded knowledge (poker-like.) While that stuff (and path finding) is the mainstay of popularized AI, its both limiting and non-interesting unless you actually care about the specific game (chess engine authors love chess, not AI.. poker bot authors love poker, not AI..)

      It is through the act of a program itself learning the knowledge necessary that makes AI both enjoyable and powerful. I can teach you a tree searching algorithm in a single day.. and I can teach you a machine learning algorithm like actor-critic in a single day.. only the later will reward you for a lifetime.

      --
      "His name was James Damore."
  8. One thing is for certain. by v(*_*)vvvv · · Score: 2

    Some think that artificial intelligence seeks to emulate the real intelligence of humans. But most of it is just software, and has little to do with real intelligence.

    There are certain problems that AI can solve, but those solutions are not "intelligent" but rather are merely "formulas" programmed by intelligent people (computer scientists).

    We get excited when these formulas emulate what a real person might do, and when we can hide the underlying machine, but that is not to say we know how people think or even how we are implemented. We are just getting better at programming.

    There are some great advancements in cognitive science, and the more we discover about how the brain works, the less it looks like it could be run by any "code". No intel inside. The brain is an organ that grows and dies, and takes its memories with it. If anything, it programs itself.

    That is not to say there haven't been advancements in AI. It too is incredibly useful.

    A good place to start:
    http://www.ted.com/search?q=brain
    http://www.ted.com/search?q=artificial+intelligence ... and wikipedia of course...

  9. Re:Stanford AI and Game Theory? by Needlzor · · Score: 5, Informative

    www.ai-class.org, by Sebastian Thrun and Peter Norvig

  10. You might try Economics instead. by meburke · · Score: 3, Insightful

    Not to belittle your choices, but this is a VERY complicated subject. My favorite introductions to game theory are, "The Compleat Strategyst" by Williams, and, "Strategy in Poker, Business and War" by McDonald. These are not trivial books, but they are easy reads into the uses of Game Theory.

    After that, you get into some Math. Read anything you can on Probability and Risk; know your Statistics and Calculus. Much of what you are looking for will be found under the subject "Decision Theory."

    I say study Economics because this is where political and economic scientific thought is making the greatest gains at this time. Game theory has a lot to do with "payoff" and Economics is a fertile field for studying payoffs. (So is Political Science, and there some good laboratories in, say, Afghanistan, Mexico and Chicago. But that's a slightly different, pragmatic, field of study.)

    My favorite definition of "politics" is: "The behavior of vying for scarce rewards." This is almost exactly a definition for Economics. At one time Economics was thought to be a sub-level of politics; it now seems the opposite is true.

    Hayak pretty much proved that economic behavior cannot be quantified because of the complexity. What is useful is deriving principles of actions under a variety of conditions to provide maximum payoffs, for the most people, under the widest variety of conditions. (An alternative course is to try to derive the largest payoffs for the fewest people under specific conditions.) AutoDesk used to have an Artificial Life laboratory that you could manipulate to learn about Genetic Algorithms and other AI behavior. Context-dependent AI can be learned through developing Neural Nets. Some of the guys I've talked to at Carnegie Mellon in the Quantitative Economics studies have warring economic artificial hybrid GA/Neural Nets, and the observations are pretty interesting.

    If it was simply a matter of rational decision making, optimum economic strategies could probably be described and tested in a much smaller AI field. However, politics and economics are burdened with mis-perceptions, human values, and stubborn beliefs. This is a big field, and you should be able to enjoy it as a hobby for the rest of your life without running into a limit of learning.

    --
    "The mind works quicker than you think!"
  11. Braitenburg and Dawkins by DanDD · · Score: 2

    First, read up on Braitenburg Vehicles and The Selfish Gene, by Richard Dawkins. Dawkins is something of a deity in the annals of evolutionary biology and is worthy of worship :-p

    Then read up on Neural Networks, start simple with a feed-forward with error backprop.

    Then try your hand at some Temporal Difference Learning.

    Then take a look at genetic algorithms, but it might help you to first understand the classic A* heuristic search algorithm. Genetic algorithms tend to be interesting search algorithms that are inspired by a genetic process, but they have little connection to the actual biological process for which they are named, so I am biased against them. This perception could just be a local cognitive minima that might be avoided with better training.

    --
    "Every time I see an adult on a bicycle, I no longer despair for the future of the human race." - H. G. Wells
    1. Re:Braitenburg and Dawkins by Rockoon · · Score: 2

      Genetic algorithms tend to be interesting search algorithms that are inspired by a genetic process, but they have little connection to the actual biological process for which they are named, so I am biased against them. This perception could just be a local cognitive minima that might be avoided with better training.

      Just stop using the name 'genetic algorithm' when thinking about them.

      Begin with a straightforward randomized state-space search method, simply remember the candidate solution with the best score. How can that be improved upon? We could keep a record of many of these randomized candidate solutions and their scores and derive new candidates using various methodologies, such as combining parts of two different candidates into a new candidate and then randomizing only minor parts of these new candidates. We can call this the 'directed randomized search method' where the performance of past candidate solutions influence the generation of new candidate solutions.

      And there you go.. the 'Directed Randomized Search Method' AKA 'Genetic Algorithm'

      What isnt so obvious is that once you start thinking about it like this, you have a greater understanding of the roles of population size, mutation rate, elitism, and so on, as well as the effects of population convergence and other observations about the population, and that the enforcement of 'generations' has no beneficial qualities (ie: it is much better to generate one new candidate at a time and moving that new information into the population immediately so that it can be leveraged immediately)

      --
      "His name was James Damore."
  12. Game Theory: A Critical Introduction by mrogers · · Score: 3
    The best technical introduction to game theory I've come across is Game Theory: A Critical Introduction by Shaun P. Hargreaves-Heap and Yanis Varoufakis, which introduces the most important concepts while placing them within their philosophical context (for example, to what extent is it reasonable to regard humans as the kind of agents assumed by game theory?). I've been studying game theory for years and wish I'd read this book a long time ago.

    If you really have no patience for philosophy, try Game Theory for Applied Economists by Robert Gibbons instead. ;-)

    John Maynard Smith's Evolution and the Theory of Games is accessible and indispensable.

    Less technical works that explore the implications of the theory in fascinating ways include The Evolution of Cooperation (the book that first got me interested in the subject) and The Complexity of Cooperation by Robert Axelrod, and anything by Brian Skyrms.

  13. Yale course "Game Theory" on Youtube by tsvk · · Score: 2

    Here is the complete Youtube playlist for the Yale course "Game Theory", lectured by Ben Polak. 24 lectures in total, about 1 h 15 min each.

    Course description: This course is an introduction to game theory and strategic thinking. Ideas such as dominance, backward induction, Nash equilibrium, evolutionary stability, commitment, credibility, asymmetric information, adverse selection, and signaling are discussed and applied to games played in class and to examples drawn from economics, politics, the movies, and elsewhere.

    I have had the intention of watching through this, but haven't had the time after the first few lectures. The material is recommended, though.

    http://www.youtube.com/playlist?list=PL6EF60E1027E1A10B

  14. Good game theory books I keep on my shelf: by tlambert · · Score: 3, Informative

    Good game theory books I keep on my shelf:

    Nonlinear Dynamics, Mathematical Biology, and Social Science (Santa Fe Institute Studies in the Sciences of Complexity Lecture Notes)
    by Joshua Epstein
    Westview Press
    ISBN: 9780201419887
    (if you know enough math for partial differential equations, this book is a must-have, since it's directly applicable to mathematically modelling open source software projects)

    The Evolution of Cooperation
    by Robert Axelrod and William D. Hamilton
    Paper: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.147.9644&rep=rep1&type=pdf
    Book: ISBN 0-465-02122-2
    Perspectives on Adaptation in Natural and Artificial Systems
    Basic Books
    ISBN: 9780195162929

    The Complexity of Cooperation: Agent-Based Models of Competition and Collaboration
    by Robert Axelrod
    Princeton University Press
    ISBN 978-0691015675

    Game Theory and the Social Contract, Vol. 1: Playing Fair
    by Ken Binmore
    MIT Press
    ISBN 978-0262023634

    Game Theory and the Social Contract, Vol. 2: Just Playing (Economic Learning and Social Evolution)
    by Ken Binmore
    MIT Press
    ISBN 978-0262024440

    Analyzing Policy: Choices, Conflicts, and Practice
    by Michael C. Munger
    W. W. Norton & Company
    ISBN 978-0393973990

    Growing Artificial Societies: Social Science from the Bottom Up (Complex Adaptive Systems
    by Joshua M. Epstein, Robert L. Axtell
    MIT Press
    ISBN 978-0262550253

    See also:

    http://www.santafe.edu/
    http://www.youtube.com/user/santafeinst

    The Brookings Institute is also active in this area (it was their math that led most of the U.S. Cold War policy and kept everyone out of a nuclear exchange with the Soviets).

    -- Terry

  15. MIT's Open CourseWare by lolococo · · Score: 2

    MIT has tons of material on AI, on their OpenCourseWare site, especially in the Electrical Engineering and Computer Science section.

  16. A simple telephone call by srussia · · Score: 2

    Tank, I need a program for AI and Game Theory... Hurry!

    --
    Set your phasers on "funky"!
  17. Consciousness by Smallpond · · Score: 3, Insightful

    I haven't seen anyone post it yet, but if your interest is in human-like intelligence, read an AI critic like Searle.

  18. Game theory by rajats · · Score: 3, Informative

    Read Avinash Dixit's Thinking Strategically to get started. It's a great book which does not use much math and can make for light reading and a great start.

  19. two MIT A.I. conferences this year by peter303 · · Score: 2

    M.I.T. had two 150th birthday conferences on A.I. this year. This would give some ideas on the state of the art and the players. Its not a systematic, pedagogical presentation.