Lecture 19: AI And Humans

# Artificial Intelligence

## AI and Humans

---

# Humans and Computers

- Creative

- Social

- Critical

- Imprecise

]

- (Mostly) deterministic

- Uncaring

- Uncritical

- Precise
]
  
---

# Example: Text Generation

<blockquote class="twitter-tweet"><p lang="en" dir="ltr">When an AI is trained on words, weird things can happen to the physical domain.<br><br>I asked the <a href="https://twitter.com/OpenAI?ref_src=twsrc%5Etfw">@OpenAI</a> API about horses.<a href="https://t.co/V6pkWxcMRC">https://t.co/V6pkWxcMRC</a> <a href="https://t.co/7errIsWq1g">pic.twitter.com/7errIsWq1g</a></p>— Janelle Shane (@JanelleCShane) <a href="https://twitter.com/JanelleCShane/status/1273296527662841856?ref_src=twsrc%5Etfw">June 17, 2020</a></blockquote>

---

# Text Generation

* The problem with text generation is that it is ill-defined

* There is no "scoring" function, nothing to "optimize"

* Instead, there are many different shades of "correct"

* Grammar might be perfect, context might be perfect, but the semantics are easy to mess up

---

# AI Dungeon

* You can play with a version of this text generator yourself!

* It is framed as a role playing game, where you play as a character and the AI is the DM

* You can "do whatever you want" (or at least what you can write), and the DM will respond

* It's a lot of fun to play with, but does not always make a lot of sense

* [AI Dungeon](https://play.aidungeon.io/)

---

# Creativity

* So, how do you judge "makes sense"?

* Or if you generate pictures, stories, movies, music, etc., how do you determine if your generator is "good"?

* It depends!

---

# Good Stories

* Let's consider the problem of generating stories

* What makes a story "good"?

* Depends on the goal!

* Possible goals: Believable characters, memorable events, uniqueness, conveying a particular message, ...

---

# Specialized measures

* The more concrete we can get with our goal, the more feasible it becomes to measure

* For example, say we want a story that conveys some information

* We could model this as a query to the mental state of the reader after reading the story

---

# QUEST

* QUEST is a cognitive model of question answering

* Given a story, a graph representation of the information contained in it is built

* Then there are operations to query this graph

* It has been shown (Cardona-Rivera et al., 2016) that this knowledge structure can predict a human reader's answers

* Of course, the model is not complete and does not pick up on all nuances

---

# Novelty

* Humans are also often drawn to novelty

* If you tell someone that a story, movie, picture, etc. was generated by a computer it may automatically be "more interesting"

* This is not necessarily bad, but it may affect the perception

---

# Social Simulations

* Social situations are complicated for AI agents

* There are several models for aspects of social behavior

* For example, one may model opinion exchange

* These models can then be validated piece by piece

---

# Self-Analysis

* The nice thing about computers it that they do what you tell them

* The problem with computers is that they do exactly what you tell them

* Who is responsible if something goes wrong?

* How can you tell why a computer did something?

---

# Literal Interpretation

---

# Explainable AI

* There is a push towards explainability

* Which inputs affected the output?

* How does the output change if the inputs change?

* This is a difficult problem for some approaches (remember Neural Networks?)

---

# Precision

* Computers have the advantage of being precise, especially with numbers

* Humans are not

* This sometimes leads to friction

---

# Precision

---

# Precision

---

# Less Precision

---

# Conclusion

---

# My Research

* I currently work on interactions between humans and AI agents

* This includes communication and collaboration

* Right now, I am also particularly interested in non-verbal communication

* For most of my research I use games!

---

# SLOTHLab

* You can find more information about my research on https://slothlab.info

* SLOTH: **S**ymbolic **L**ogic and **O**perations for **T**ransactions with **H**umans

* My dissertation was about using Dynamic Epistemic Logic to model belief updates in games

* This included One Night Ultimate Werewolf, a game which involves lying!

---

# Hanabi

---

# Pandemic

---

# Planning and Plan Recognition

* The general problem we are trying to solve with this is cooperation

* The AI agent is trying to plan how to reach a goal (e.g. win the game)

* But they need to take into account what the other (human) player is trying to do

* The process which we use for this is called *Plan Recognition*

---

# Non-Verbal communication

* People communicate in a variety of ways

* We have looked into timing, gaze and facial expressions

* Timing: Longer thinking times may indicate uncertainty

* Gaze: Where a player looks indicates what they are planning or considering

* Facial Expressions: Happyness/Anger/Surprise may give information about hidden information

---

# Timing in Hanabi

* We developed an improved version of the Hanabi agent that measures how long the human player thinks

* This information can be used to determine how certain they are about what to do

* The agent uses this information to help with interpreting the human player's actions

* We performed an online experiment to show that this can actually help

---

# Timing in Hearthstone

---

# Gaze In Hanabi

---

# Adaptive Game using Facial Expression

* Another Master's student, Jonathan, is working on a procedurally generated game

* The idea is that the game changes depending on how the player feels about it

* If the player is "too angry" about the game state, it becomes nicer

* If the player is "too happy", it becomes more challenging

---

# Dixit

---

# Thank you

* This was the first time I taught AI

* I hope you enjoyed the topic, and my way of teaching it

* Please participate in the evaluation!

---

# The Future

* If you are interested in working on projects in AI and games, let me know

* I will be working at [CalPoly Pomona](https://www.cpp.edu)

* If you are interested in a Master's thesis, email me

* If you are considering a PhD about AI in games (especially in the US), email me

---

# Reminder: Schedule

* Next week, we will do another review for the final exam

* The final presentations of the projects are scheduled for July 16th, 6pm

* You also have to submit a short (2-3 pages) report about what you did

* The final exam is officially scheduled for July 9th

* You will have the entire week (Monday - Friday) for the exam again

---

# Exam

* Topics: Everything since exam 1 (Lecture 9 - 19, Planning - AI and Humans)

* You can use the slides, videos, the book, etc. during the exam

* Ask **me** if anything is unclear, not your classmates

* The answers have to be **your own**

---

# References

* [Previewing OpenAI's new API,](https://aiweirdness.com/post/621186154843324416/all-your-questions-answered)

* [AI Dungeon](https://play.aidungeon.io/)

* [Question Answering in the Context of Stories Generated by Computers](http://www.rogel.io/content/2-publications/cardona-rivera2016question.pdf)

* [Opinionated Virtual Characters](https://www.aaai.org/ojs/index.php/AIIDE/article/download/5232/5088)

* [Hanabi Online](http://hanabi.slothlab.info)