GECCO 2011

Evolving Patch-

based Terrains for use in Video Games

William Raffe, Fabio Zambetta, Xiaodong LiRMIT University, Melbourne, Australia

Evolving Patch-based TerrainsGECCO’11, July 14th

Dublin, Ireland

W. Raffe, F. Zambetta, X. Li

RMIT University, Australia

Outline

� Background� Procedural Terrain Generation.

� Using EA in Content Generation.

� Approach� Patch-based Terrain Generation

• Extracting patches, re-combining patches, and terrain generation parameters.

� Evolutionary Algorithm• Genetic representation and operators.• Parent and Gene Selection.

� Results � Sample run.

� Capabilities in generating terrain for games.

� Future Work


Dublin, Ireland



Procedural Terrain

Generation

� The creation of virtual terrain through algorithmic means.� Reduces the time and skill to create terrain.

� Focus on game maps� Reduce production costs.

� Increase replay value.

� Popular techniques include:� Fractals

� Erosion simulation

� Noise generators

� Delaunay and Voronoi diagrams


Dublin, Ireland



Height-map Terrains


Dublin, Ireland



Why use EA?

� Iterative evolutionary process allows for maps to be refined over time.� Feedback from users or fitness functions direct the terrain

generation process.

� More control • Generate-and-test vs Constructive.

� Evolutionary algorithms successfully used on other types of game content.� Euphoria character animation

� Search-based Procedural Content Generation (SBPCG). � EA popular option for SBPCG.

� Has been used to change game content for each player based on their individual preferences and how they play the game.


Dublin, Ireland



Patch-based Approach

� Extract patches from sample terrains.

� Initial population can either be:� The sample terrains themselves.OR

� Constructed of randomly selected patches.

� Conduct Two-Level interactive evolution.� Parent Selection

� Gene Selection

� Genetic operators create children by swapping individual patches of the selected parents.

� Repeat interactive evolution on new generation.


Dublin, Ireland



Extracting Patches

� Program takes 8 sample terrains.

� Sample terrains are divided into uniform sized patches.� Extra height-map data on all sides of each patch is

extracted to allow for overlapping.

� Each patch is:� Stored as a 2D array of height values (i.e. a small

height-map).� Placed in a global patch database.� Given a unique ID that is referenced when

constructing new terrains.


Dublin, Ireland



Recombining Patches

(Roof-Tiling)

� Choose a starting patch


Dublin, Ireland



Recombining Patches

(Roof-Tiling)


� Choose the next patch and overlap it with the one before it.


Dublin, Ireland



Recombining Patches

(Roof-Tiling)



� Keep adding patches until an entire row is complete.


Dublin, Ireland



Recombining Patches

(Roof-Tiling)




� Repeat this to create a second row.


Dublin, Ireland



Recombining Patches

(Roof-Tiling)





� Overlap the second row with the first.


Dublin, Ireland



Recombining Patches

(Roof-Tiling)





� Overlap the second row with the first.

� Repeat the process of creating rows and adding them to the row before until a full terrain is made.


Dublin, Ireland



Stitching Patches

� Patches are overlapped to ensure a smooth transition from one patch to the next.

� Height values in the overlap region are interpolated between the two patches.� The height of a vertex is

weighted towards one patch or the other based on how far through the overlap region it is.

� Cubic Spline Interpolation is used to smooth the entries into the overlap region.


Dublin, Ireland



Terrain Parameters

� Number of Patches:� Less Patches: Larger patch size. Ensures smooth

flow in terrain but does not allow for variety.

� More Patches: Smaller patch size. Allows for control over finer details but higher chance jagged terrain.

� Overlap size:� Large Overlap: Can blend patches together to make

rolling terrain but lose feature details in each patch.

� Small Overlap: Creates hard edges which are good for cliff faces but makes the boarders of each patch more obvious.


Dublin, Ireland



Patch and Overlap Size

Reduce

Overlap

Size

Increase

number of

patchesNumber of Patches = 4 (2x2)Overlap Size = 80%

Number of Patches = 81 (9x9)

Overlap Size = 20%


Dublin, Ireland



EA - Representation

� Genotype Representation: Patch Map� A 2D array of patch ID numbers (a Patch Map). Each patch

ID corresponds to a patch in the current dataset of patches (extracted from sample terrains at start-up).

� Dimensions of array reflect how many patches are used to construct each terrain.

13 455

22

34 18

15 9

40

Terrain made up of 9 patches (3x3)


Dublin, Ireland



EA - Crossover� Uniform Crossover:

� One parent chosen as base parent.

� Each patch given a randomly generated probability to crossover

� If this value is less than the Crossover Rate parameter, patch is switched for patch in same position on other parent, otherwise the patch is kept where it is.

� Crossover Rate between 0 and 1:� Due to randomly chosen base parent, a rate over 0.5 is equivalent

to that under 0.5.

� We mostly used a crossover rate of 0.5 – patches from both parents have equal chance of being present in each child.

Crossover rate of 0.2. First parent (white) is base parent.


Dublin, Ireland



Mutation

� Each patch given probability to mutate and is compared to the Mutation Rate parameter, same as crossover.� If the patch is to be mutated, in is replaced with a

randomly chosen patch from the dataset of all existing patches.

� Mutation Rate between 0 and 1:� All of our experiments used a mutation rate between 0.1 or

0.3.

� <0.1 results in not enough variety in children.

� >0.3 results in too much difference between parent and child and frequent undesirable mutations.


Dublin, Ireland



Interactive Evolution:

Parent Selection

� User picks terrains that exhibit the features they desire.

� User chooses 0 – 2 parents.� No terrains other than those chosen by the user are

considered for breeding.

� 0 parents: All offspring are randomly generated.

� 1 parent: All offspring are mutated versions of the parent.

� 2 parents: All offspring are the result of crossover between parents and individual mutations.

� Elitist selection: User selected parents each have one child that is an exact copy.


Dublin, Ireland




Parent Selection


Dublin, Ireland




Gene Selection

� User specifies which patches are immune to crossover and mutation.� All selected patches will not change from the base

parent.

� All unselected patches may experience crossover an mutation as usual.

� Inverse mutation rate: � More patches selected, higher chance of mutation for

unselected patches.

� High chance of mutation when only a few patches unselected.


Dublin, Ireland




Gene Selection


Dublin, Ireland



Why use Gene Selection?

� Reduce user fatigue!

� In short:� Protect good patches from mutation.

� Promote the change of bad patches.

Swap out these two patches andkeep everything else as is. Difficultwhen only using Parent Selection.


Dublin, Ireland



Results: Sample Run (1)

� Generation 1: � Randomly generated.


Dublin, Ireland




� Generation 4: � Main feature starts to appear.


Dublin, Ireland




� Generation 6: � Prominent main feature but a few wrong patches.


Dublin, Ireland




� Four results from four separate runs

� All with the same user goal

9 generations 7 generations

13 generations 16 generations


Dublin, Ireland



Results: Game Terrain (1)

� Original map from the game Halo (left).

� Terrain generated by our system (right).� Generated through multiple runs with different parameter settings

� The resulting terrain from one run was used as a sample terrain to the next run.


Dublin, Ireland



Results: Game Terrain (2)

� Variations on the result terrain.� Small variations can result in new experiences for the

player without needing to learn an entirely new map.


Dublin, Ireland



Future Challenges

� Move towards automated fitness evaluation.� Retain some manner of Gene Selection.

� Refine type of terrain being generated by focusing on a single game genre.� Currently, we generate Virtual Terrain but we want

one system to generate complete Game Maps

� Investigate player profiling to drive the evolutionary process.� Allows for possibilities in tailoring game maps to

individual players.


Dublin, Ireland



Questions?

GECCO 2011

Education

Transcript of GECCO 2011