Meta-Reasoning

2019-07-24_08h24_18.png

Mistakes in Reasoning and Learning

Reasoning

2019-07-24_08h30_37.png

We easily got stuck unable to progress without undoing earlier moves. Our reasoning erred.

Learning

Learning by correcting mistakes (if agent had used explanation based learning)

Task of Meta cognition here: How did I learn this knowledge where I committed the mistake? Explanation based learning. So how do I fix that in Explanation based learning.

Knowledge gaps

There could also be gaps, which meta cognition may aim to resolve.

mc_1.gif

Reasoning gaps

In case of reasoning gap, for example in below case of means-ends analysis, MC may split goals in to smaller goals (problem reduction) and then solve.

2019-07-24_08h41_44.png

Blurred line between MC and Deliberation

2019-07-24_08h45_51.png

  • MC and Deliberation are not distinctly separated as depicted above.
  • They could be thought of as overalapping boxes, and it is ok, if we are unable to say what task comes under Deliberation or MC.
  • Whats more important is how to solve the problem at hand.

Strategy Selection

MC has number of strategies to select. Which one to select?

mc_2.gif

It may select based on largely 3 criteria, and whichever in that is most important,

  • What related knowledge does it have specific to the problem? For ex, if for a problem, the cases are not avialbale, then CBR is not a choice
  • Computational efficiency some choices may be more efficient. If a problem is very close to a case in past then CBR might be a good choice.
  • Quality of solution Some comes with a guarantee (a quality) of solutions. For eg, logic.

Similarly for learning methods,

2019-07-24_08h59_18.png

  • Some may be applicable to the problem, and some may not. For exm, if examples come, one at a time, then ICP.
  • Computational efficiency and Quality of solutions like earlier.

Strategy Integration

  • The agent need not stick to one strategy fully. It may shift on the go as needed.
  • In CBR, during adaptation, it may find rule based reasoning as most optimal, and so on. Note also how MC provides needed inputs below.

mc_3.gif

  • In MEA, when we hit cul de sac, we used problem reduction which is another method to resolve and then again used MEA for those sub goals. Classic example it is.

Process of Meta-Reasoning

MC may use one of the same reasoning strategies on itself, to choose strategies.

Meta Meta Reasoning

No need caz MC can reason over itself.

mc_4.gif

Goal Based Anatomy

Complicated. So given as video.

In [1]:
from IPython.display import YouTubeVideo
YouTubeVideo('f7PXyfIcV2M')
Out[1]:

Connections

In [2]:
from IPython.display import YouTubeVideo
YouTubeVideo('CytsqYxmDjQ')
Out[2]:

Meta-Reasoning in CS7637

This is gold. Watch it.

In [3]:
from IPython.display import YouTubeVideo
YouTubeVideo('fBHwgGNVm_8')
Out[3]: