Self-Modelling Group: Meeting 23/05/2024
===

###### tags: `insight` `pretending` `learning`

:::info
- **Presenter:** Zoe, Limits of self modeling ([slides](https://docs.google.com/presentation/d/1kYviKYKmAueWsHP_F_HJZj51tId8h9IC_9pTQbkIBQY/edit?usp=sharing))
- **Date:** May 23, 2024 15:30 PM (OXFORD)
- **Participants:**
- Matan
- Zoe
- Noam
- Maya
- Nicole
- dates in a jar
- **Notes:** Noam

:::

## Intro:

● No direct approach to all information; there is a limit, compression, and a sense of abstraction. Understanding the limit will help to understand its structure. The boundaries will reveal the shape.

● Pretending behavior shown in figures. The same pattern between pretending and not pretending, with a longer general RT in the pretense condition.

● Aha! as a sudden moment of expansion; 'losing' the bias for a moment.

● Local insight, within a given game, and global insight, understanding the role of symmetry.

● Methodological note: Choosing between options, this is a process we have access to. Battleships show that, but this is after the options have already come to mind. Here we are interested in an earlier phase. Can people have the idea of symmetry coming to mind? Perhaps not, and this could be revealing the limits of self-modeling.

● Using our advanced machinery equipped with abilities to view behavior and not just numbers, with respect to seeing participants actually playing the game.

## Results:

● Could be an intriguing interaction with a need for symmetry—'I need to see the white square also on the right side'; this could influence behavior and make a participant with high OCD tendencies less ideal.

● A cool index—the difference between symmetrical squares. We would expect that insight would cause the second click (parallel square) to become shorter. Therefore the difference would become larger as the learning increases.

● People seem to report insight only in the last game—therefore, it could be beneficial to add one or two more boards.

● Game 5 could be a very clear symmetrical variation (Matan’s note; might need additional explanation here).

- [name=Matan ] If we make the pattern relatively hard to discren in boards 1-4, then easy to discren in board 5, and hard again in board 6, we can compare the board-5 effect in pretend and non-pretend.

## Pretending condition:

● Compare someone who played with someone who did not play at all and see if there is different pretend behavior.
*

● Manipulate whether pretending is first or starting with actually playing and measure its effect.

- one question is whether pretenders assume that an insight moment will occur
- the other question is whether pretenders can replicate/emulate the dynamics/timepoint of how an insight occurs
- we need to be careful to make sure that pretenders do not have reason to assume that players were given the information about symmetry. E.g. "Players are presented with the following boards(show revealed boards) in which a hidden rule determines the location of black squares. They are not told about this rule and need to reveal each black square."

● To have a line that shows the orientation of the symmetry: horizontal, diagonal. Check when participants realize this.

● Pressing 'J' before making a decision as a proxy for insight.
- [name=Matan]
> "Once you press J, you get additional 5 free points. But this comes at a cost: from the moment you play J until you finish the board, each white square you touch will cost you 5 points instead of 1."

## Postscript by Matan: potential ways to quantify learning

In the following, we define $x'$ as the "twin" of square $x$ on the other side of the board. Let's imagine that two sets of squares were revealed: $W$ (revealed to be white) and $B$ (revealed to be black). $W'$ is the group of *unrevealed* squares whose twin has been shown to be white, and $B'$ is the group of unrevealed squares whose twin has been shown to be black.

### Board-wise measures of learning:

- Total number of points earned.
- Time spent on board in seconds.
- The ratio between these two numbers.
- Number of $W'$ clicks out of all white clicks.
- RT($B$)-RT($B'$) (where $B$ is the first black square to be revealed for each pair, $B'$ is the second, and RT is the response time *from the previous click (whether it is the twin or not)*.

### Manipulations to test explicit learning:

- "Once you press J, you get additional 5 free points. But this comes at a cost: from the moment you press J until you finish the board, each white square you touch will cost you 5 points instead of 1."
- [name=Matan] Just to say that I personally like this idea a lot.
* A board in which one half of 2 shapes is fully uncovered, e.g there are no black squares apart from $B'+B$.
We could also introduce this at the end of the study in case there are dependencies, to distinguish people who have discovered the symmetry from people who have not. (It gives people a bit of a hint if done during the study, but that might be interesting too).

### Click-wise measures of learning:
- Slow accurate clicks where number of $W'$ clicks before the click is higher than the number of $W'$ clicks after the click (within and across boards). Not sure how to quantify this exacly though.

### Measures that don't directly indicate learning
- board-wise transition probability from left half to right half (what is the probability that my next click will be in the other side of the board?)
- Hover-log motion energy along the x axis vs y axis.