Intro

Agent will not learn concepts, but connections between existing concepts

2019-07-02_17h37_51.png

Why is above easy for humans, hard for robots?

Example Retrieving a Cup

2019-07-02_17h43_14.png

Concept Space

2019-07-02_17h45_01.png

Prior Knowledge

2019-07-02_18h22_13.png

Constructing for a brick..

2019-07-02_17h48_38.png

  • Note Causaulity is also captured. because in sentence captured as yellow lines in prior knowledge.
  • Note not all elements participate in causality. Example: A brick is simply heavy in above example.

Abstraction

2019-07-02_17h51_44.png

  • Bowl is abstracted to object
  • Note only causal explanations are abstracted.

Transfer

Now how to prove an object is a cup? We have this concept space to start, and prior knowledge as we just saw.

2019-07-02_18h00_35.png

Bottom up

First using Bowl abstraction, we could build below construct. 2019-07-02_18h03_41.png 2019-07-02_18h01_30.png

Next from briefcase, similar one..

2019-07-02_18h14_05.png

2019-07-02_18h15_11.png

Same way for Glass but see how it extends from briefcase and bowl..

2019-07-02_18h15_50.png

2019-07-02_18h16_52.png

Note how one's effect has become another's cause. Bowl's effect becomes Glass's one of causes.

Next comes brick

2019-07-02_18h18_39.png

2019-07-02_18h19_29.png

Now finally for a cup

2019-07-02_18h23_56.png

So,

2019-07-02_18h24_23.png

Summary Animation

anim

  • Note once above representation is constructed proof is backwards. First goal starts with Object is a cup, and then as problem reduction, reduced to 2 sub goals. Its stabld and enables drinking. Next we have to prove that, and so on. This also involves planning with sub goals as pre conditions The end point causes become post conditions, and then they are pre condition for new sets and so on.

  • Another example: Get me that important document from last Tuesday Agent should prove any document it finds is important or not from past with sub goals. For example from prior knowledge it might know which ever document has a signature is important (cause and effect), then agent coudl work on that..

  • Papers are blowing off from my desk. How can I stop that? In prior knowledge, something which has certain properties eventually stopped blowing off paper. May be a paper weight. Now you see a cup. A cup could have similar cause-effect property, acting as a weight. So it could be used to stop the paper blowing off.

Exercise

Prove an object is an instance of a Mug.

2019-07-02_18h39_55.png

Answer is No because unable to link heat transfer and heat protection from prior knowledge

2019-07-02_18h50_39.png

Process

  • Start with what we want to prove. Here, a mug. Break down in to sub goals.
  • Try proving each sub goal from prior knowledge. Except heat protection, other sub goals we have already seen in last example of cup how could we prove.
  • For heat transfer we could not link wiht pot because no prior knowledge that links them via causal relation

Exercise 2

An Oven Mitt fits here providing that missing link. Also wooden spoon fits bypassing pot.

2019-07-02_18h55_46.png

2019-07-02_18h56_06.png