6 Matching Annotations
  1. Aug 2016
    1. Covered in this paper

      • Transformation invariance (includes scale and translational)


    1. According to their model, our feelings affect behavior along a continuum between, on one end, something they term “generativeness” (that is, how likely you are to explore something that may end up having a good result, if doing so involves risk) and, on the other, “defensiveness” (when you are focussed on avoiding negative outcomes, forgoing opportunities in the process). It’s a concept akin to what the Columbia University psychologist Tory Higgins calls promotion and prevention—that is, the decision to work toward something or to direct your energy toward avoiding something else. When we are constantly monitoring our behavior, we tend to be on guard and act defensively. We tend to prevent rather than to promote.
    1. tldr: Problem: agents should have a guess of where unseen enemies are.<br> 3 Approaches:

      1. Occupancy Maps. Simplest method. When the enemy slips out of view, you are 100% sure of his location. A second later you are about 98% sure, but he could also be in surround areas. After a little while you're like 10% sure that he's in any given location.
      2. Particle Filters. Since occupancy maps can be computationally expensive, you only keep track of the most likely places.
      3. Simulacra. Calculate an actual simulated agent that acts like a real agent (to the best of your knowledge), just don't render them. This will probably get you the best, most accurate results.
    2. The occupancy map and particle filter techniques described above both make implicit use of a simulacrum

      Explicit simulcrum solves the problem that the previous two techniques shared: that agents aren't equally likely to walk in every direction.

    3. Occupancy maps have several weaknesses.

      Good quiz question

    1. producesaprobabilityofdetectionthatisbasedon thebrightnessand sizeofthetarget, aswellasthebrightnessofthetarget'sbackground.

      Uses the difference between foreground and background brightness to determine visibility. (A ninja dressed in black is hard to see against a black wall.) The problem arises when you have a bright wall in front of the ninja and the algorithm thinks the contrast makes the ninja more visible. This is solved by the GBBA method.