YouTube

Got a YouTube account?

New: enable viewer-created translations and captions on your YouTube channel!

English subtitles

← abm 9 6 wrapup+future

Get Embed Code
5 Languages

Showing Revision 4 created 03/17/2017 by seonaid.

  1. So essentially, we've now reached the end
    of the course. But before I leave you,
  2. I wanted to speculate on three trends and
    necessities of agent-based modelling
  3. that I see are vitally important as we
    move forward with agent-based modelling.
  4. And this is my particular take, these are
    things I'm passionately interested in,
  5. but it's also capitalizing on trends we're
    seeing in other areas of computational
  6. sciences, especially computational social
    sciences and biological-socio-ecological
  7. sciences. And one of them is the automatic
    generation of agent rules. Now, this is
  8. something I talked about in 9.1, right?
    We have this huge amount of big data
  9. that's now being collected about what
    humans and entities are doing around the
  10. world, right? We have sensors on
    everything, we have the internet of things
  11. exploding left and right. We have social
    data, we have app data, we have data about
  12. the phones in everybody's pockets, right?
    And the question is, can we use that
  13. in some way to define rules automatically
    that capture the essence of human actions
  14. in those spaces, right? Like define how
    humans move through the world, define how
  15. humans interact with the internet of
    things, right? Or define how the internet
  16. of things interacts with itself, right,
    as cars become more and more automated.
  17. Of course, these rules also need to be
    validated, right? And one of the ways we
  18. can do the validation is by building up
    these rules, making predictions
  19. with them seeing these predictions
    carried out.
  20. Now why would it be interesting to do this
    at an agent level, right?
  21. If we have all this data, why can't we
    just use the data, right?
  22. But without the agent rules that predict
    individual-level behavior,
  23. we don't really have the ability to assess
    what would happen if we were to change
  24. the incentive structure for one particular
    individual in that space.
  25. Now causal state modelling gives us one
    example of this.
  26. But we could use many others,
    you could use decision trees,
  27. you could use the associative rules, right.
  28. You could use classifier systems. There are
    a lot of methods
  29. out there that allow you to do this.
  30. And with all these sources of data, big data,
  31. administrative data is not new, but the ability to process it in large amounts is,
  32. natural language data, text to speech data,
  33. social data, app data, really being able to
  34. capitalize on this will powerfully change the way
  35. agent-based modelling is perceived in the world around us, right.
  36. And give us a much more powerful tool kit
    to address some of those situations.
  37. Now, of course, one of the big problems with some of this
  38. data and one of the problems that people
    often complain about, right is that
  39. trace data is essentially just digital exhaust.
  40. It is data that doesn't tell you anything.
  41. So we need ways to validate those models
  42. against real-world data and calibrate
  43. those models in order to show that they're
    actually working well, right.
  44. We need rigorous guidelines, I believe, to follow,
  45. to show that our models, the agent-based
    models, have been validated appropriately.
  46. I often think of this as a statistics-like
    suite of tests, right.
  47. Statistics is very good. The discipline of
    statistics is very good.
  48. at saying, "If you have data looks like
    this and your outputs look like this,
  49. then these are the various tests
    that you need to apply," right.
  50. And in some cases, all we would be doing is
  51. basically discovering which of those
    appropriate statistical tests, that had
  52. already existed, would be most likely to
    apply in that particular situation, right.
  53. And then, what that means is that lends
    a lot of credence and credibility to
  54. the idea that this model I created and
    run through these tests to compare to
  55. empirical data that I've seen have now then, have helped me to validate and
  56. increase my confidence in this model.
  57. Now, of course, being able to calibrate our models
  58. in order to increase that level of validation
    would also be useful.
  59. So making tools like BehaviorSearch, which
    we discussed earlier, easier to use
  60. so users can calibrate models automatically against
  61. real-world data would be a very powerful step forward in this space.
  62. Finally, and this kind of stems off of
    both of those other thoughts about
  63. the future, if we could build models
    automatically... if we could construct
  64. them from these vast amounts of data that
  65. are coming in - and we could continuously
    validate them in some automatic way
  66. where we have a suite of tests that
    we know will tell us whether or not
  67. the model is behaving accurately, then we
  68. theoretically could build a tool which
    will automatically construct a model
  69. on the basis of streaming data, right?
  70. and by that I mean it will be pulling
    down... for instance...
  71. stock-ticker data, Twitter data, whatever
  72. it needs, and continually making a
    model of, say, the socioeconomic status
  73. of the country as a whole, for instance,
    or something like that.
  74. Or maybe it's pulling in data from the
    internet of things, that looks at
  75. sensors that are attached to all of the
    street lights in a particular city
  76. that detects the traffic flow, and then
    making continual predictions
  77. about whether or not there's going to be
  78. traffic jams in certain areas
  79. so that policy makers and city managers
    can, in real time, change signage
  80. in order to have people move around
    traffic jams.
  81. If that was the case... if we had this
    ability, this tool to build these
  82. frameworks in real time that were
    predictive, well, not predictive,
  83. but able to forecast and explore scenarios
    into the future,
  84. we could then use this to support
    real-time decision-making.
  85. So those are three areas - the use of
    streaming data, big data
  86. and the ability to automatically create
    rules from big data,
  87. and this validation and calibration
    and improvement, that I think are
  88. critical for the future of agent-based
    modelling.
  89. Like I said, that's it. Unit 9, we talked
    about big data ABM, we talked about
  90. design guidelines. We talked about
    the uses of ABM for communication
  91. and education. We talked about advanced
    programming constructs like map and
  92. reduce and run and run-result task.
  93. We talked about participatory simulations,
    systems dynamic modelling
  94. then we talked about extensions and the
    future of agent-based modelling.
  95. You'll see the Unit 9 slides, and you'll
    see the Unit 9 tests coming up shortly.
  96. But I want to thank everyone for
    participating in this course.
  97. I hope you enjoyed it. I enjoyed
    teaching it for sure,
  98. and I've really enjoyed a lot of
  99. the discussions that we've been having
  100. in the forums, and through the YouTube
    office hours, and things like that.
  101. And I hope that you are able to gain
    something out of this.
  102. Please, stay in touch as you go forward
  103. and use agent-based modelling in your
  104. own work or for your own fun,
    for that matter.