Lecture 14: Terrain Generation

# AI in Digital Entertainment

### Terrain Generation

---

# Terrains

---

# Terrain Generation

* Say we have a 3D game where you visit a variety of islands

* We want to generate each island procedurally

* How do we do that?

---

# Terrain Representation

* For now, let's ignore caves

* Each island is a 2D surface in 3D space

* We can "just" store (and generate) the height value for each point

* However, that will still be way too many values, so we use a grid instead

---

# Grid-based Heightmap

---

# Bilinear Interpolation Artifacts

---

# Bicubic Interpolation

---

# Bicubic Interpolation

---

# Spline Interpolation

---

# Noise-Based Terrain Generation

---

# Noise-Based Terrain Generation

* What is noise? Random numbers (more or less)

* So, let's generate a terrain by picking height values at random

* What could possibly go wrong?

---

# Random Terrain

---

# Difference-Based Terrain Generation

* When we just pick random numbers, adjacent grid cells may have very different values

* This will lead to a very rugged looking terrain

* What would be more realistic? Don't let the height values *change* between adjacent vertices arbitrarily

* Instead, use the existing values of the neighbors, and *add* a small difference value

---

# Difference-Based Terrain

---

# Perlin Noise Terrain Generation

* The terrain already looks a lot better, and we may control how rugged it looks by controlling how much the difference can be at most

* It could be smoother, though!

* Idea: Instead of controlling the difference between heights, we control the difference of the difference in height, or rather: We generate a smooth derivative

* Perlin Noise uses this idea!

---

# Perlin Noise Terrain

---

# Agent-Based Terrain Generation

---

# Agent-Based Terrain Generation

* Instead of generating the entire terrain at once, it may be worthwhile to divide the task into smaller subtasks

* We can then implement an agent for each subtask

* What is the general process of creating landscapes?
   - Coastline generation 
   - Landform generation 
   - Erosion
   
---

# Coastline Generation

* Coastline agents draw the outline of the landscape

* Generation starts with one agent that splits until each agent is responsible for a small enough part of the map

* Within this part, the agent determines random points, using attractors and repulsor points for the coast line

* The agents work concurrently on different parts of the map

* When the agents finish, smoothing agents wander around and change the elevation of nodes according to their neighbors

---

# Coastlines

---

# Beaches and Mountains

* Beach agents wander around the parts of the map that are close to sea level

* They may create beaches by flattening the land

* Likewise, mountain agents wander the parts of the map that are above a certain height level

* When they find such spots, they may increase the height to create mountains

* They will wander in a certain direction to create a ridge, with a chance to turn

* After beaches and mountains are created, the landscape is smoothed again

---

# Beaches and Mountains

---

# Rivers

* After beaches and mountains are created, it is time to make rivers

* For each river, a start-point in the mountains, and an end point on the coast are determined

* An agent starts at the end-point and tries to wander uphill towards the start-point

* If they succeed, they then go back downhill and dig a bed for the river

---

# Rivers

---

# Agent-Based Terrain Generation

* Each agent has several parameters that determine how it generates "its" feature

* By varying these parameters, different types of landscapes can be generated

* For example, the mountain agent creates ridges, but turns with a certain probability. Changing this probability controls how straight mountain-ridges will be

* The number of agents of each type also dictates how prevalent each feature is

---

# Cellular Automata-Based Terrain Generation

---

# Cellular Automata

* Cellular Automata are zero-player games on a grid 
  
  * The basic idea is: Every cell has some states (sometimes just "on" and "off"), and there are rules for how to update the grid 
  
  * Most famous Cellular Automaton rules: Conway's Game of Life
  
     - If a cell is "on" and has two or three (of its eight) neighbors that are also "on", it will continue being "on", otherwise it is turned "off"
     - If a cell is "off" and has exactly three (of its eight) neighbors that are "on", it will turn "on", otherwise it stays turned "off"
  
---

# Cellular Automata

---

# Cellular Automata

* How is this useful?
  
  * Different rules make different shapes!
  
  * Let's say we want to generate caves:
      - Fill the grid randomly with "on" and "off" cells 
      - A cell is turned "on" if five or more of its neighbors are "on"
      - Run for a number of generations
      
---

# Cellular Automata

---

# Cellular Automata Terrain Generation

* We can also interpret our states as "lower" and "raise" terrain 
  
  * The rules determine how much raised/lowered neighbors influence a cell to be raised/lowered 
  
  * The terrain output also depends on how many generations we run 
  
  * Alternatively: Each state determines the "type" of terrain of a larger map fragment (beach, mountain, plains, etc.)
  
---

# Cellular Automata Terrain

---

# Next Week

* Project Presentations

* Keep your presentation to about 15-20 minutes

---

# Conclusion

---

# Conclusion

- It was the first time I taught this class, I enjoyed it a lot!
  
  - I hope you enjoyed taking it, too
  
  - Y'all were a great audience!
  
  - I hope you all had fun with your projects 
  
  - If you are interested in pursuing research around game AI (in Costa Rica or elsewhere), come talk to me!
  
---

# My Research

- I am currently working on non-verbal communication in video games 
  
  - For example, how can an AI agent use facial expression analysis to understand a human player better 
  
  - We're also looking at the timing of actions, and eye tracking
  
  - Currently our applications are cooperative games, such as Hanabi or Pandemic 
  
  - I'm interested in AI for games that plays with human players
  
  - If you're interested in working with me, come talk to me!

---

# My previous research: One Night Ultimate Werewolf

---

# My research: Hanabi

---

# My student's research: Pandemic

---

# Feedback

- If there is an official evaluation, please participate
  
  - You can also let me know if there is anything you particularly liked/disliked (for example by posting anonymously and/or privately on Piazza)
  
  - What was your favorite/least favorite part of the class (if you want to share)?

---

##  Thank you all for your participation!

---

# References

* Chapter 4 of [Procedural Content Generation](http://pcgbook.com/) by Noor Shaker, Julian Togelius, and Mark J. Nelson (free PDF available)
  
  * [Controlled Procedural Terrain Generation Using Software Agents](http://larc.unt.edu/ian/pubs/terrain.pdf), by Jonathon Doran and Ian Parberry
  
  * [Cave Generation with Cellular Automata](https://www.raywenderlich.com/2425-procedural-level-generation-in-games-using-a-cellular-automaton-part-1)
  
  * [Algorithmic Terrain with Cellular Automata](https://demonstrations.wolfram.com/AlgorithmicTerrainWithCellularAutomata/) Wolfram Alpha Demonstration Project

* [Generating Terrain with Cellular Automata](http://ideabyre.com/world-building/erosion/erosion_model)