Mercurial > cortex

annotate org/intro.org @ 22:157b416152ea

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
continuing splitting
author Robert McIntyre <rlm@mit.edu>
date Sun, 23 Oct 2011 23:35:04 -0700
parents
children e965675ec4d0
rev   line source
rlm@22 1 #+title: Simulated Senses
rlm@22 2 #+author: Robert McIntyre
rlm@22 3 #+email: rlm@mit.edu
rlm@22 4 #+description: Simulating senses for AI research using JMonkeyEngine3
rlm@22 5 #+SETUPFILE: ../../aurellem/org/setup.org
rlm@22 6 #+INCLUDE: ../../aurellem/org/level-0.org
rlm@22 7 #+babel: :mkdirp yes :noweb yes
rlm@22 8
rlm@22 9 * Background
rlm@22 10 Artificial Intelligence has tried and failed for more than half a
rlm@22 11 century to produce programs as flexible, creative, and "intelligent"
rlm@22 12 as the human mind itself. Clearly, we are still missing some important
rlm@22 13 ideas concerning intelligent programs or we would have strong AI
rlm@22 14 already. What idea could be missing?
rlm@22 15
rlm@22 16 When Turing first proposed his famous "Turing Test" in the
rlm@22 17 groundbreaking paper [[./sources/turing.pdf][/Computing Machines and Intelligence/]], he gave
rlm@22 18 little importance to how a computer program might interact with the
rlm@22 19 world:
rlm@22 20
rlm@22 21 #+BEGIN_QUOTE
rlm@22 22 \ldquo{}We need not be too concerned about the legs, eyes, etc. The example of
rlm@22 23 Miss Helen Keller shows that education can take place provided that
rlm@22 24 communication in both directions between teacher and pupil can take
rlm@22 25 place by some means or other.\rdquo{}
rlm@22 26 #+END_QUOTE
rlm@22 27
rlm@22 28 And from the example of Hellen Keller he went on to assume that the
rlm@22 29 only thing a fledgling AI program could need by way of communication
rlm@22 30 is a teletypewriter. But Hellen Keller did possess vision and hearing
rlm@22 31 for the first few months of her life, and her tactile sense was far
rlm@22 32 more rich than any text-stream could hope to achieve. She possessed a
rlm@22 33 body she could move freely, and had continual access to the real world
rlm@22 34 to learn from her actions.
rlm@22 35
rlm@22 36 I believe that our programs are suffering from too little sensory
rlm@22 37 input to become really intelligent. Imagine for a moment that you
rlm@22 38 lived in a world completely cut off form all sensory stimulation. You
rlm@22 39 have no eyes to see, no ears to hear, no mouth to speak. No body, no
rlm@22 40 taste, no feeling whatsoever. The only sense you get at all is a
rlm@22 41 single point of light, flickering on and off in the void. If this was
rlm@22 42 your life from birth, you would never learn anything, and could never
rlm@22 43 become intelligent. Actual humans placed in sensory deprivation
rlm@22 44 chambers experience hallucinations and can begin to loose their sense
rlm@22 45 of reality in as little as 15 minutes[sensory-deprivation]. Most of
rlm@22 46 the time, the programs we write are in exactly this situation. They do
rlm@22 47 not interface with cameras and microphones, and they do not control a
rlm@22 48 real or simulated body or interact with any sort of world.
rlm@22 49
rlm@22 50
rlm@22 51 * Simulation vs. Reality
rlm@22 52 I want demonstrate that multiple senses are what enable
rlm@22 53 intelligence. There are two ways of playing around with senses and
rlm@22 54 computer programs:
rlm@22 55
rlm@22 56 The first is to go entirely with simulation: virtual world, virtual
rlm@22 57 character, virtual senses. The advantages are that when everything is
rlm@22 58 a simulation, experiments in that simulation are absolutely
rlm@22 59 reproducible. It's also easier to change the character and world to
rlm@22 60 explore new situations and different sensory combinations.
rlm@22 61
rlm@22 62
rlm@22 63 ** Issues with Simulation
rlm@22 64
rlm@22 65 If the world is to be simulated on a computer, then not only do you
rlm@22 66 have to worry about whether the character's senses are rich enough to
rlm@22 67 learn from the world, but whether the world itself is rendered with
rlm@22 68 enough detail and realism to give enough working material to the
rlm@22 69 character's senses. To name just a few difficulties facing modern
rlm@22 70 physics simulators: destructibility of the environment, simulation of
rlm@22 71 water/other fluids, large areas, nonrigid bodies, lots of objects,
rlm@22 72 smoke. I don't know of any computer simulation that would allow a
rlm@22 73 character to take a rock and grind it into fine dust, then use that
rlm@22 74 dust to make a clay sculpture, at least not without spending years
rlm@22 75 calculating the interactions of every single small grain of
rlm@22 76 dust. Maybe a simulated world with today's limitations doesn't provide
rlm@22 77 enough richness for real intelligence to evolve.
rlm@22 78
rlm@22 79 ** Issues with Reality
rlm@22 80
rlm@22 81 The other approach for playing with senses is to hook your software up
rlm@22 82 to real cameras, microphones, robots, etc., and let it loose in the
rlm@22 83 real world. This has the advantage of eliminating concerns about
rlm@22 84 simulating the world at the expense of increasing the complexity of
rlm@22 85 implementing the senses. Instead of just grabbing the current rendered
rlm@22 86 frame for processing, you have to use an actual camera with real
rlm@22 87 lenses and interact with photons to get an image. It is much harder to
rlm@22 88 change the character, which is now partly a physical robot of some
rlm@22 89 sort, since doing so involves changing things around in the real world
rlm@22 90 instead of modifying lines of code. While the real world is very rich
rlm@22 91 and definitely provides enough stimulation for intelligence to develop
rlm@22 92 as evidenced by our own existence, it is also uncontrollable in the
rlm@22 93 sense that a particular situation cannot be recreated perfectly or
rlm@22 94 saved for later use. It is harder to conduct science because it is
rlm@22 95 harder to repeat an experiment. The worst thing about using the real
rlm@22 96 world instead of a simulation is the matter of time. Instead of
rlm@22 97 simulated time you get the constant and unstoppable flow of real
rlm@22 98 time. This severely limits the sorts of software you can use to
rlm@22 99 program the AI because all sense inputs must be handled in real
rlm@22 100 time. Complicated ideas may have to be implemented in hardware or may
rlm@22 101 simply be impossible given the current speed of our
rlm@22 102 processors. Contrast this with a simulation, in which the flow of time
rlm@22 103 in the simulated world can be slowed down to accommodate the
rlm@22 104 limitations of the character's programming. In terms of cost, doing
rlm@22 105 everything in software is far cheaper than building custom real-time
rlm@22 106 hardware. All you need is a laptop and some patience.
rlm@22 107
rlm@22 108 * Choose a Simulation Engine
rlm@22 109
rlm@22 110 Mainly because of issues with controlling the flow of time, I chose to
rlm@22 111 simulate both the world and the character. I set out to make a minimal
rlm@22 112 world in which I could embed a character with multiple senses. My main
rlm@22 113 goal is to make an environment where I can perform further experiments
rlm@22 114 in simulated senses.
rlm@22 115
rlm@22 116 As Carl Sagan once said, "If you wish to make an apple pie from
rlm@22 117 scratch, you must first invent the universe." I examined many
rlm@22 118 different 3D environments to try and find something I would use as the
rlm@22 119 base for my simulation; eventually the choice came down to three
rlm@22 120 engines: the Quake II engine, the Source Engine, and jMonkeyEngine.
rlm@22 121
rlm@22 122 ** Quake II/Jake2
rlm@22 123
rlm@22 124 I spent a bit more than a month working with the Quake II Engine from
rlm@22 125 ID software to see if I could use it for my purposes. All the source
rlm@22 126 code was released by ID software into the Public Domain several years
rlm@22 127 ago, and as a result it has been ported and modified for many
rlm@22 128 different reasons. This engine was famous for its advanced use of
rlm@22 129 realistic shading and had decent and fast physics
rlm@22 130 simulation. Researchers at Princeton [[http://www.nature.com/nature/journal/v461/n7266/pdf/nature08499.pdf][used this code]] to study spatial
rlm@22 131 information encoding in the hippocampal cells of rats. Those
rlm@22 132 researchers created a special Quake II level that simulated a maze,
rlm@22 133 and added an interface where a mouse could run around inside a ball in
rlm@22 134 various directions to move the character in the simulated maze. They
rlm@22 135 measured hippocampal activity during this exercise to try and tease
rlm@22 136 out the method in which spatial data was stored in that area of the
rlm@22 137 brain. I find this promising because if a real living rat can interact
rlm@22 138 with a computer simulation of a maze in the same way as it interacts
rlm@22 139 with a real-world maze, then maybe that simulation is close enough to
rlm@22 140 reality that a simulated sense of vision and motor control interacting
rlm@22 141 with that simulation could reveal useful information about the real
rlm@22 142 thing. It happens that there is a Java port of the original C source
rlm@22 143 code called Jake2. The port demonstrates Java's OpenGL bindings and
rlm@22 144 runs anywhere from 90% to 105% as fast as the C version. After
rlm@22 145 reviewing much of the source of Jake2, I eventually rejected it
rlm@22 146 because the engine is too tied to the concept of a first-person
rlm@22 147 shooter game. One of the problems I had was that there does not seem
rlm@22 148 to be any easy way to attach multiple cameras to a single
rlm@22 149 character. There are also several physics clipping issues that are
rlm@22 150 corrected in a way that only applies to the main character and does
rlm@22 151 not apply to arbitrary objects. While there is a large community of
rlm@22 152 level modders, I couldn't find a community to support using the engine
rlm@22 153 to make new things.
rlm@22 154
rlm@22 155 ** Source Engine
rlm@22 156
rlm@22 157 The Source Engine evolved from the Quake II and Quake I engines and is
rlm@22 158 used by Valve in the Half-Life series of games. The physics simulation
rlm@22 159 in the Source Engine is quite accurate and probably the best out of
rlm@22 160 all the engines I investigated. There is also an extensive community
rlm@22 161 actively working with the engine. However, applications that use the
rlm@22 162 Source Engine must be written in C++, the code is not open, it only
rlm@22 163 runs on Windows, and the tools that come with the SDK to handle models
rlm@22 164 and textures are complicated and awkward to use.
rlm@22 165
rlm@22 166 ** jMonkeyEngine
rlm@22 167
rlm@22 168 jMonkeyEngine is a new library for creating games in Java. It uses
rlm@22 169 OpenGL to render to the screen and uses screengraphs to avoid drawing
rlm@22 170 things that do not appear on the screen. It has an active community
rlm@22 171 and several games in the pipeline. The engine was not built to serve
rlm@22 172 any particular game but is instead meant to be used for any 3D
rlm@22 173 game. After experimenting with each of these three engines and a few
rlm@22 174 others for about 2 months I settled on jMonkeyEngine. I chose it
rlm@22 175 because it had the most features out of all the open projects I looked
rlm@22 176 at, and because I could then write my code in Clojure, an
rlm@22 177 implementation of LISP that runs on the JVM...