The Late Late Show with Andrei Alexandrescu

From: Andrei Alexandrescu (andrei_at_cs.washington.edu)
Date: Mon Oct 20 2003 - 13:18:20 PDT

  • Next message: Karthik Gopalratnam: "Review 2 - Doyle & Patil - "Two Theses of Knowledge Representation""

    <html>
    <pre>

    Evolving Robot Tank Controllers by Jacob Eisenstein

    Summary: The paper evaluates the design and evolution of a Robocode
    controller conceived using genetic programming techniques. The
    controller is written in an language amenable to genetic
    transformations, TableRex.

    Most important ideas:

    * Genetic programming can create successful controllers. This sounds
    like a tautology, but given that the whole Genetic Programming field
    is in need for identity and confirmation, the point is important.
    Eisenstein's contribution was to find a suitable encoding (TableRex)
    and transformations to that encoding that yielded increasingly good
    behavior.

    * Evaluation and definition of the fitness function influence genetic
    code evolution in essential ways. Eisenstein showed that the way we
    humans distinguish a "good" robot from a "bad" one has effect on the
    behavior of robots. Too coarse of a measuring method, and that's not
    enough for the algorithm to learn. His algorithms yielded controllers
    that did well by some measure of performance, without being too "good"
    at an intuitive level.

    Largest flaws:

    * Overspecialization is not intelligence. Using the real adversaries
    as sparring partners is a luxury often not present in the real world.
    One could think that new, improved robots would do more of the same
    things that existing robots do, but there's absolutely no evidence
    from the paper that the genetic controller could do well in situations
    it hasn't been trained for. Actually, there is negative evidence in
    that only changing as minor a detail as starting positions degrades
    performance to unimpressive.

    * Each event handling function is independent from the others. There
    is much opportunity for speeding up learning if the functions can
    exploit the obvious correlations that exist between events.

    Open questions:

    * How would an expressive, general purpose genetic programming
    language look like? Researchers have used LISP S-expressions in the
    past, and Eisenstein uses a limited (no loops!) linear language. The
    ideal language suitable for meaningful genetic transformations is
    still around the corner.

    * How to build reflexes, goals, and instincts into genetic evolution?
    In the real world, learning babies do have many reflexes and instincts
    built in; there are many mysteries about infant and child behavior
    that escape scientists. How can we incorporate the ineffable mistery
    that makes life so wonderfully fervid, into the initial condition and
    entire learning process of genetic algorithms?

    </pre>
    </html>


  • Next message: Karthik Gopalratnam: "Review 2 - Doyle & Patil - "Two Theses of Knowledge Representation""

    This archive was generated by hypermail 2.1.6 : Mon Oct 20 2003 - 13:18:01 PDT