SlideShare une entreprise Scribd logo
1  sur  46
Télécharger pour lire hors ligne
Discover the world at Leiden UniversityDiscover the world at Leiden University
The Role of
Evolutionary Computation
in Game AI
Mike Preuss @ S3 ACM GECCO summer school, Praha
July 10, 2019
Discover the world at Leiden University
● my roots in Evolutionary Computation
● game AI today: a forming field
● Procedural Content Generation
● Balancing
● GVGAI/competitions/rolling horizon EA
roadmap
2
Discover the world at Leiden University
Evolutionary Computation and me
●most likely the last PhD student of Hans-Paul
Schwefel (co-inventor of the Evolution Strategy)
●most important message: expect the unexpected
●focus on multi-modal optimization (detect
several good, different solutions simultaneously)
●a bit on multi-objective optimization
●some work on experimental methodology
3
Discover the world at Leiden University
PhD (2013) and book (2015)
most important results:
●the advantage of niching (if any) can only come from
coordinated placement
●in contrast: SLS is randomized placement
●simulations/theory: unfortunately, we cannot speed
up finding the global optimum
●but if we go for coverage (find most if not all good
optima) we can gain a lot
●invented NEA2 as top algorithm (2015), lesson
learned: for niching, also use the objective values
modern name for this stuff: quality diversity
4
Discover the world at Leiden University
GECCO multimodal optimization competition
• builds on a fixed selection of 20 problems made
from 12 functions in different dimensions
• averages over 3 different rankings:
- average PR (peak rate) values
- static F1 (takes the amount of points in the
answer set into account)
- dynamic F1 (same but as integral over time
for the whole runtime)
• only global optima are sought
• the number of optima is low for most problems
http://www.epitropakis.co.uk/gecco2019/
5
Discover the world at Leiden University
what we did not expect so early
6
Discover the world at Leiden University
bla
a forming field
a forming field
Discover the world at Leiden University
what about
game AI?
current game AI is:
• playing well
• generating content
• modeling players
see 2018 book:
Artificial Intelligence
and Games, by
Yannakakis/Togelius
http://gameaibook.org/book.pdf
video: fully generated
game worlds
in No Man’s Sky (2016)
8
Discover the world at Leiden University
when game AI becomes… artificially intelligent
ongoing discussion in game AI: how intelligent is current game AI?
see recent gamasutra post:
http://www.gamasutra.com/view/news/253974/When_artificial_intelligence_in_video_games_becomesartificially_intellige
nt.php
"My prediction is a little bit controversial.
I think the next giant leap of game AI is actually artificial intelligence."
- Alex J. Champandard at GDC 2013
Ken de Jong (at GECCO 2017, Berlin) on the question:
What do you see as dominant research direction in Evolutionary Computation for
the next years?
“It is time to apply our methods to real problems now.”
9
Discover the world at Leiden University
classical AI techniques in games
even in the early 2000s, lots of standard techniques in use (and developers showing
little interest for a change):
• finite state machines (FSM)
• hierarchical FSM
• scripting
basically mostly used for Non-Player Character (NPC) control, also:
• cellular automata
• influence maps
• basic optimization approaches
• game theory
(still useful for the right problem, but we have so much more now)
10
Discover the world at Leiden University
modern approach
• we treat all aspects of games, now also game design/production
• we use what fits best for a specific problem
everything is an agent:
AI director, opponent AI controller, content generator, ....
game AI
specific algorithms
classic AI /
machine learning
in-game AI use
computational intelligence
game design/production
11
Discover the world at Leiden University
game ai overall view
• game AI traditionally encompasses mostly NPC behavior and bots
• control happens during the game, design support before deployment
• design support is rather new, but a highly dynamic research area
game AI
design supportcontrol
directors
strategic (bots)
/supportive AI
PCGNPC behavior balancing
NPC = non-player character, PCG = procedural content generation
12
Discover the world at Leiden University
panoramic view
game AI components, player centric view, from:
Yannakakis, G.N.; Togelius, J., "A Panorama of Artificial and Computational Intelligence in Games," in IEEE
TCIAIG, 2014
13
Discover the world at Leiden University
why apply CI methods to games?
contrary to board games,
• game trees often not applicable
• incomplete information
• concurrency: during planning phase, game situation
changes
• quantifying a game situation is not trivial
➔ good and fast approximations are needed
evolutionary optimization is
• versatile, flexible, still works (somehow)
• copes with noise and strange search spaces
• can be asked to deliver a result at any time
14
Discover the world at Leiden University
Evolutionary and AI
•not fully up to date, but clearly shows relations
•boundaries much more fuzzy nowadays
15
Discover the world at Leiden University
CI and games: a field forming
• 1999: Blondie24, learning checkers with CI
• 2005 first Computational Intelligence in Games (CIG) conference
• IEEE TCIAIG Journal (Transactions on CI and Artificial
Intelligence in Games) since 2009
• EvoGames track in Evo*, DETA track in GECCO since 2009/2011
• many “neighbor” conferences, etc. AIIDE, FDG, gameai/nucl.ai
conf. (not strictly CI, but CI welcome)
• general approach target oriented, not technique oriented
• 2012: first Dagstuhl seminar on AI and CI in Games
• 2015: second Dagstuhl seminar (integration)
• 2017: third Dagstuhl seminar (AI-supported design)
• 2018/2019: journal and conference renamed to IEEE Transactions
on Games and IEEE Conference on Games
• 2019: first Shonan seminar on AI for Games
16
Discover the world at Leiden University
one small step for [a] man….
what recently happened:
17
AlphaGo: Go
(MCTS, expert knowledge, self-play)
AlphaGoZero: Go
(MCTS-, self-play)
Deep Reinforcement Learning: ALE
(pixel input, self-play)
AlphaStar: StarCraft
(MCTS?, pixel input, self-play, expert knowledge)
AlphaZero: Go, Chess, Shogi
(MCTS-, self-play)
Go-Explore: Montezuma
(population based exploration)
Obstacle Tower Challenge
(population based training)
ForTheWin: Capture the Flag
(2 speed rnns, population based
learning, population based training)
Discover the world at Leiden University
combining EC and Deep Reinforcement Learning
•some (limited) successes in weight optimization with EC
•population-based learning:
- GANs, multiple agent architectures
- resembles co-evolutionary approaches, same problems (how to control)
•population-based based training:
- generating test cases as needed
- procedural content generation (PCG), mainly done via EC methods
•not really exploited yet: multi-modal/multi-objective approaches
•maybe more things to transfer?
18
Discover the world at Leiden University
what I did in Game AI
●procedural generation of strategy game maps
● organizing StarCraft I competitions, survey
● some works on balancing (also using EC):
○there are usually many solutions in balancing
○the problem is very multi-modal
●recently: facet orchestration (how to combine
several generators, e.g. visuals, audio)
●very recently: transfer AlphaAnything to new
problems (not from the games domain)
●understand how it works on small games
19
Discover the world at Leiden University
Obstacle Tower Challenge
current DRL challenge
for artificial general
intelligence (AGI)
• levels increase in
difficulty
• levels procedurally
generated
• decorations change
• includes minigames
(Sokoban)
runs in 2 phases:
- qualification (gone)
- finals (up to July
15th, 2019)
https://www.aicrowd.com/challenges/unity-
obstacle-tower-challenge
20
Discover the world at Leiden University
bla
Procedural Content Generation
Discover the world at Leiden University
search-based vs. generate-and-test
22
Discover the world at Leiden University
Evolutionary Algorithms
• initial solutions needed: existing or randomly generated solutions
• fitness function shall measure the targeted properties (designer input)
• matching variation operators need to be established
23
Discover the world at Leiden University
why using EAs in PCG?
pros:
- flexible, only needs good utility function
- not very sensitive concerning noise/mistakes
- anytime method: we can obtain a result whenever needed
cons:
- optimality cannot be guaranteed (except for very few special cases)
- stagnation phases of unpredictable length (break?)
- depends on matching representation and operators
- slow on simple problems
24
Discover the world at Leiden University
back in 2009: galactic arms race
25
Discover the world at Leiden University
how are the new weapons produced?
• very small population, 3 solutions (candidates, individuals)
• interactive selection: player removes existing solution when new one is generated
• weapons are created via a parametrized particle system by complex neural
network
• cgNEAT: neural network topology optimized by means of evolutionary algorithm
26
Discover the world at Leiden University
a sample gallery
27
Discover the world at Leiden University
StarCraft I map evolution (2009-2012)
28
Discover the world at Leiden University
RTS map design with Sentient Sketchbook
Preuss, Liapis, Togelius. Searching for Good and Diverse Game Levels. IEEE CIG 2014.
29
Discover the world at Leiden University
gamesbyangelina and the procjam
• Michael Cook’s AI that is designed to intelligently design video games:
http://www.gamesbyangelina.org/
• PROCJAM: yearly game jam with strong emphasis on PCG
http://itch.io/jam/procjam
• coming soon: facet orchestration
30
Discover the world at Leiden University
Project seminar
nd
balancing
Discover the world at Leiden University
Game AI hot topics
• procedural content generation (PCG)
- large game worlds are expensive to make
- represent specific (designer) styles
• believability of non-player character (NPC) behavior
- more humanlike behavior (also evoke and show emotions)
- better cooperation of units (team AI)
- human/AI interaction
• personalization
- preference modeling (what do they like?), player type analysis
- dynamic adaptation of game content and mechanics
balancing is
everywhere
32
Discover the world at Leiden University
balancing project with BlueByte
• prototype from BlueByte
• ZombieVillage game, related to tower defense
• several parameters: can we balance automatically?
33
Discover the world at Leiden University
balancing project with BlueByte
• number of (most important) parameters reduced to 5
• interchange between manual and automated (integrated) balancing
• core result: both steps needed, many different solutions exist
Marlene Beyer, Aleksandr Agureikin, Alexander Anokhin, Christoph Laenger, Felix Nolte, Jonas Winterberg, Marcel Renka,
Martin Rieger, Nicolas Pflanzl, Mike Preuss, Vanessa Volz: An integrated process for game balancing. CIG 2016
34
Discover the world at Leiden University
tackling a complex RTS
Mike Preuss, Thomas Pfeiffer, Nicolas Pflanzl, Vanessa Volz: Integrated Balancing of an RTS Game:
Case Study and Toolbox Refinement. CIG 2018
35
Discover the world at Leiden University
bla
nd
general game AI/competitions
Discover the world at Leiden University
can we also have general (game) AI?
• instead of this:
• we rather want this:
game 1
game 2
game 3AI 3
AI 2
AI 1
develop for
develop for
develop for
game 1
game 2
game 3
AI 1
learn to play
learn to play
learn to play
37
Discover the world at Leiden University
GVGAI ENVIRONMENT
• general video-game playing (GVGP) succeeds general game-playing
• targeted at more complex, non-deterministic games
• language first shaped during Dagstuhl seminar CI and AI in games 2012
• we want to obtain controllers that can play “any game”
• important publications:
John Levine, Clare B. Congdon, Michal Bída, Marc Ebner, Graham Kendall, Simon Lucas, Risto Miikkulainen,
Tom Schaul, and Tommy Thompson. General Video Game Playing. Dagstuhl Follow-up, 6:1–7, 2013.
Tom Schaul. A Video Game Description Language for Model-based or Interactive Learning. In Proceedings of the
IEEE Conference on Computational Intelligence in Games, Niagara Falls, 2013. IEEE Press
http://www.gvgai.net
38
Discover the world at Leiden University
GVGAI competition
• build general AI controllers
• learn to play many games, not one
• many different challenges
• many successful controllers
use Monte-Carlo Tree search
• but Rolling Horizon EA also an option
• but how can this work?
• “traditional track”: 1-player planning
• employs forward model (action simulation, may be non-deterministic)
• new: 2-player planning, learning (no forward model)
• now also: level generation, rule generation
39
Discover the world at Leiden University
some example games
Infection:
Missile Command:
Pacman
40
Discover the world at Leiden University
VGDL – game and level definitions
41
Discover the world at Leiden University
A SAMPLE (RANDOM) CONTROLLER
42
Discover the world at Leiden University
Rolling Horizon EA for search based planning
• basic idea: evolve whole sequence of moves at once, apply only first action
• hybrids with MCTS like playouts are also possible
• simple extension with shift buffer (only remove latest action) performs better
Raluca Gaina, Simon Lucas, Diego Perez. Rolling Horizon Evolution Enhancements in General Video
Game Playing. IEEE CIG 2017
43
Discover the world at Leiden University
current competitions held at COG 2019
• 4th Angry Birds Level Generation Competition
• Bot Bowl I
• Fighting Game AI Competition
• First TextWorld Problems: A Reinforcement and Language Learning Challenge
• Geometry Friends Game AI Competition
• General Video Game AI Competitions (Learning Track)
• Hanabi Competition
• Hearthstone AI competition
• MicroRTS AI Competition
• Short Video Competition
• StarCraft AI Competition
• Strategy Card Game AI Competition
lots of demo and explanation videos included:
http://ieee-cog.org/competitions_conference/
44
Discover the world at Leiden University
conclusions
•EC is part of AI and has always been
•Procedural Content Generation getting very important for games
industry but also for Deep Learning
•evolutionary approaches have great potential for improving other
algorithms, e.g. RHEA (rolling horizon EA) instead of MCTS
•3 important applications of EC algorithms in game AI:
PCG, balancing, search-based planning
45
Discover the world at Leiden UniversityDiscover the world at Leiden University
Don’t
panic!
46

Contenu connexe

Similaire à The Role of Evolutionary Computation in Game AI

A survey on AI in computer games
A survey on AI in computer gamesA survey on AI in computer games
A survey on AI in computer games
RedwanIslam12
 

Similaire à The Role of Evolutionary Computation in Game AI (20)

SAIG Overview March 2011
SAIG Overview March 2011SAIG Overview March 2011
SAIG Overview March 2011
 
Jacob Sherson (University of Aarhus) - Science@home
Jacob Sherson (University of Aarhus) - Science@homeJacob Sherson (University of Aarhus) - Science@home
Jacob Sherson (University of Aarhus) - Science@home
 
Matching Game Mechanics and Human Computation Tasks in Games with a Purpose
Matching Game Mechanics and Human Computation Tasks in Games with a PurposeMatching Game Mechanics and Human Computation Tasks in Games with a Purpose
Matching Game Mechanics and Human Computation Tasks in Games with a Purpose
 
Matching Game Mechanics and Human Computation Tasks in Games with a Purpose
Matching Game Mechanics and Human Computation Tasks in Games with a PurposeMatching Game Mechanics and Human Computation Tasks in Games with a Purpose
Matching Game Mechanics and Human Computation Tasks in Games with a Purpose
 
How to do science in a large IT company (ICPC World Finals 2021, Moscow)
How to do science in a large IT company (ICPC World Finals 2021, Moscow)How to do science in a large IT company (ICPC World Finals 2021, Moscow)
How to do science in a large IT company (ICPC World Finals 2021, Moscow)
 
Deep Learning, an interactive introduction for NLP-ers
Deep Learning, an interactive introduction for NLP-ersDeep Learning, an interactive introduction for NLP-ers
Deep Learning, an interactive introduction for NLP-ers
 
earning by s/doing/h4ck1ng/ - Our experience learning application security th...
earning by s/doing/h4ck1ng/ - Our experience learning application security th...earning by s/doing/h4ck1ng/ - Our experience learning application security th...
earning by s/doing/h4ck1ng/ - Our experience learning application security th...
 
Science Game Lab
Science Game LabScience Game Lab
Science Game Lab
 
AI Based Game Design - Teaching how to expand designers' artistic palette wit...
AI Based Game Design - Teaching how to expand designers' artistic palette wit...AI Based Game Design - Teaching how to expand designers' artistic palette wit...
AI Based Game Design - Teaching how to expand designers' artistic palette wit...
 
A survey on AI in computer games
A survey on AI in computer gamesA survey on AI in computer games
A survey on AI in computer games
 
Science and Videogames. Computational intelligence in videogames
Science and Videogames. Computational intelligence in videogamesScience and Videogames. Computational intelligence in videogames
Science and Videogames. Computational intelligence in videogames
 
Open ai openpower
Open ai openpowerOpen ai openpower
Open ai openpower
 
MILA DL & RL summer school highlights
MILA DL & RL summer school highlights MILA DL & RL summer school highlights
MILA DL & RL summer school highlights
 
5th world feb2006_smu_v1.0
5th world feb2006_smu_v1.05th world feb2006_smu_v1.0
5th world feb2006_smu_v1.0
 
Global Game Jam Overview
Global Game Jam OverviewGlobal Game Jam Overview
Global Game Jam Overview
 
Interactive Video Search: Where is the User in the Age of Deep Learning?
Interactive Video Search: Where is the User in the Age of Deep Learning?Interactive Video Search: Where is the User in the Age of Deep Learning?
Interactive Video Search: Where is the User in the Age of Deep Learning?
 
Our shelter just collapsed, who didn’t calculate correctly
Our shelter just collapsed, who didn’t calculate correctlyOur shelter just collapsed, who didn’t calculate correctly
Our shelter just collapsed, who didn’t calculate correctly
 
Teaching AI through retro gaming
Teaching AI through retro gamingTeaching AI through retro gaming
Teaching AI through retro gaming
 
Artificial Intelligence in Gaming
Artificial Intelligence in GamingArtificial Intelligence in Gaming
Artificial Intelligence in Gaming
 
AGI: Still relevant?
AGI: Still relevant?AGI: Still relevant?
AGI: Still relevant?
 

Dernier

Activity 01 - Artificial Culture (1).pdf
Activity 01 - Artificial Culture (1).pdfActivity 01 - Artificial Culture (1).pdf
Activity 01 - Artificial Culture (1).pdf
ciinovamais
 
Seal of Good Local Governance (SGLG) 2024Final.pptx
Seal of Good Local Governance (SGLG) 2024Final.pptxSeal of Good Local Governance (SGLG) 2024Final.pptx
Seal of Good Local Governance (SGLG) 2024Final.pptx
negromaestrong
 
Russian Escort Service in Delhi 11k Hotel Foreigner Russian Call Girls in Delhi
Russian Escort Service in Delhi 11k Hotel Foreigner Russian Call Girls in DelhiRussian Escort Service in Delhi 11k Hotel Foreigner Russian Call Girls in Delhi
Russian Escort Service in Delhi 11k Hotel Foreigner Russian Call Girls in Delhi
kauryashika82
 

Dernier (20)

Application orientated numerical on hev.ppt
Application orientated numerical on hev.pptApplication orientated numerical on hev.ppt
Application orientated numerical on hev.ppt
 
UGC NET Paper 1 Mathematical Reasoning & Aptitude.pdf
UGC NET Paper 1 Mathematical Reasoning & Aptitude.pdfUGC NET Paper 1 Mathematical Reasoning & Aptitude.pdf
UGC NET Paper 1 Mathematical Reasoning & Aptitude.pdf
 
SKILL OF INTRODUCING THE LESSON MICRO SKILLS.pptx
SKILL OF INTRODUCING THE LESSON MICRO SKILLS.pptxSKILL OF INTRODUCING THE LESSON MICRO SKILLS.pptx
SKILL OF INTRODUCING THE LESSON MICRO SKILLS.pptx
 
How to Give a Domain for a Field in Odoo 17
How to Give a Domain for a Field in Odoo 17How to Give a Domain for a Field in Odoo 17
How to Give a Domain for a Field in Odoo 17
 
Grant Readiness 101 TechSoup and Remy Consulting
Grant Readiness 101 TechSoup and Remy ConsultingGrant Readiness 101 TechSoup and Remy Consulting
Grant Readiness 101 TechSoup and Remy Consulting
 
Understanding Accommodations and Modifications
Understanding  Accommodations and ModificationsUnderstanding  Accommodations and Modifications
Understanding Accommodations and Modifications
 
Activity 01 - Artificial Culture (1).pdf
Activity 01 - Artificial Culture (1).pdfActivity 01 - Artificial Culture (1).pdf
Activity 01 - Artificial Culture (1).pdf
 
2024-NATIONAL-LEARNING-CAMP-AND-OTHER.pptx
2024-NATIONAL-LEARNING-CAMP-AND-OTHER.pptx2024-NATIONAL-LEARNING-CAMP-AND-OTHER.pptx
2024-NATIONAL-LEARNING-CAMP-AND-OTHER.pptx
 
Sociology 101 Demonstration of Learning Exhibit
Sociology 101 Demonstration of Learning ExhibitSociology 101 Demonstration of Learning Exhibit
Sociology 101 Demonstration of Learning Exhibit
 
Seal of Good Local Governance (SGLG) 2024Final.pptx
Seal of Good Local Governance (SGLG) 2024Final.pptxSeal of Good Local Governance (SGLG) 2024Final.pptx
Seal of Good Local Governance (SGLG) 2024Final.pptx
 
Micro-Scholarship, What it is, How can it help me.pdf
Micro-Scholarship, What it is, How can it help me.pdfMicro-Scholarship, What it is, How can it help me.pdf
Micro-Scholarship, What it is, How can it help me.pdf
 
Third Battle of Panipat detailed notes.pptx
Third Battle of Panipat detailed notes.pptxThird Battle of Panipat detailed notes.pptx
Third Battle of Panipat detailed notes.pptx
 
On National Teacher Day, meet the 2024-25 Kenan Fellows
On National Teacher Day, meet the 2024-25 Kenan FellowsOn National Teacher Day, meet the 2024-25 Kenan Fellows
On National Teacher Day, meet the 2024-25 Kenan Fellows
 
Explore beautiful and ugly buildings. Mathematics helps us create beautiful d...
Explore beautiful and ugly buildings. Mathematics helps us create beautiful d...Explore beautiful and ugly buildings. Mathematics helps us create beautiful d...
Explore beautiful and ugly buildings. Mathematics helps us create beautiful d...
 
Russian Escort Service in Delhi 11k Hotel Foreigner Russian Call Girls in Delhi
Russian Escort Service in Delhi 11k Hotel Foreigner Russian Call Girls in DelhiRussian Escort Service in Delhi 11k Hotel Foreigner Russian Call Girls in Delhi
Russian Escort Service in Delhi 11k Hotel Foreigner Russian Call Girls in Delhi
 
ICT role in 21st century education and it's challenges.
ICT role in 21st century education and it's challenges.ICT role in 21st century education and it's challenges.
ICT role in 21st century education and it's challenges.
 
Dyslexia AI Workshop for Slideshare.pptx
Dyslexia AI Workshop for Slideshare.pptxDyslexia AI Workshop for Slideshare.pptx
Dyslexia AI Workshop for Slideshare.pptx
 
Introduction to Nonprofit Accounting: The Basics
Introduction to Nonprofit Accounting: The BasicsIntroduction to Nonprofit Accounting: The Basics
Introduction to Nonprofit Accounting: The Basics
 
This PowerPoint helps students to consider the concept of infinity.
This PowerPoint helps students to consider the concept of infinity.This PowerPoint helps students to consider the concept of infinity.
This PowerPoint helps students to consider the concept of infinity.
 
TỔNG ÔN TẬP THI VÀO LỚP 10 MÔN TIẾNG ANH NĂM HỌC 2023 - 2024 CÓ ĐÁP ÁN (NGỮ Â...
TỔNG ÔN TẬP THI VÀO LỚP 10 MÔN TIẾNG ANH NĂM HỌC 2023 - 2024 CÓ ĐÁP ÁN (NGỮ Â...TỔNG ÔN TẬP THI VÀO LỚP 10 MÔN TIẾNG ANH NĂM HỌC 2023 - 2024 CÓ ĐÁP ÁN (NGỮ Â...
TỔNG ÔN TẬP THI VÀO LỚP 10 MÔN TIẾNG ANH NĂM HỌC 2023 - 2024 CÓ ĐÁP ÁN (NGỮ Â...
 

The Role of Evolutionary Computation in Game AI

  • 1. Discover the world at Leiden UniversityDiscover the world at Leiden University The Role of Evolutionary Computation in Game AI Mike Preuss @ S3 ACM GECCO summer school, Praha July 10, 2019
  • 2. Discover the world at Leiden University ● my roots in Evolutionary Computation ● game AI today: a forming field ● Procedural Content Generation ● Balancing ● GVGAI/competitions/rolling horizon EA roadmap 2
  • 3. Discover the world at Leiden University Evolutionary Computation and me ●most likely the last PhD student of Hans-Paul Schwefel (co-inventor of the Evolution Strategy) ●most important message: expect the unexpected ●focus on multi-modal optimization (detect several good, different solutions simultaneously) ●a bit on multi-objective optimization ●some work on experimental methodology 3
  • 4. Discover the world at Leiden University PhD (2013) and book (2015) most important results: ●the advantage of niching (if any) can only come from coordinated placement ●in contrast: SLS is randomized placement ●simulations/theory: unfortunately, we cannot speed up finding the global optimum ●but if we go for coverage (find most if not all good optima) we can gain a lot ●invented NEA2 as top algorithm (2015), lesson learned: for niching, also use the objective values modern name for this stuff: quality diversity 4
  • 5. Discover the world at Leiden University GECCO multimodal optimization competition • builds on a fixed selection of 20 problems made from 12 functions in different dimensions • averages over 3 different rankings: - average PR (peak rate) values - static F1 (takes the amount of points in the answer set into account) - dynamic F1 (same but as integral over time for the whole runtime) • only global optima are sought • the number of optima is low for most problems http://www.epitropakis.co.uk/gecco2019/ 5
  • 6. Discover the world at Leiden University what we did not expect so early 6
  • 7. Discover the world at Leiden University bla a forming field a forming field
  • 8. Discover the world at Leiden University what about game AI? current game AI is: • playing well • generating content • modeling players see 2018 book: Artificial Intelligence and Games, by Yannakakis/Togelius http://gameaibook.org/book.pdf video: fully generated game worlds in No Man’s Sky (2016) 8
  • 9. Discover the world at Leiden University when game AI becomes… artificially intelligent ongoing discussion in game AI: how intelligent is current game AI? see recent gamasutra post: http://www.gamasutra.com/view/news/253974/When_artificial_intelligence_in_video_games_becomesartificially_intellige nt.php "My prediction is a little bit controversial. I think the next giant leap of game AI is actually artificial intelligence." - Alex J. Champandard at GDC 2013 Ken de Jong (at GECCO 2017, Berlin) on the question: What do you see as dominant research direction in Evolutionary Computation for the next years? “It is time to apply our methods to real problems now.” 9
  • 10. Discover the world at Leiden University classical AI techniques in games even in the early 2000s, lots of standard techniques in use (and developers showing little interest for a change): • finite state machines (FSM) • hierarchical FSM • scripting basically mostly used for Non-Player Character (NPC) control, also: • cellular automata • influence maps • basic optimization approaches • game theory (still useful for the right problem, but we have so much more now) 10
  • 11. Discover the world at Leiden University modern approach • we treat all aspects of games, now also game design/production • we use what fits best for a specific problem everything is an agent: AI director, opponent AI controller, content generator, .... game AI specific algorithms classic AI / machine learning in-game AI use computational intelligence game design/production 11
  • 12. Discover the world at Leiden University game ai overall view • game AI traditionally encompasses mostly NPC behavior and bots • control happens during the game, design support before deployment • design support is rather new, but a highly dynamic research area game AI design supportcontrol directors strategic (bots) /supportive AI PCGNPC behavior balancing NPC = non-player character, PCG = procedural content generation 12
  • 13. Discover the world at Leiden University panoramic view game AI components, player centric view, from: Yannakakis, G.N.; Togelius, J., "A Panorama of Artificial and Computational Intelligence in Games," in IEEE TCIAIG, 2014 13
  • 14. Discover the world at Leiden University why apply CI methods to games? contrary to board games, • game trees often not applicable • incomplete information • concurrency: during planning phase, game situation changes • quantifying a game situation is not trivial ➔ good and fast approximations are needed evolutionary optimization is • versatile, flexible, still works (somehow) • copes with noise and strange search spaces • can be asked to deliver a result at any time 14
  • 15. Discover the world at Leiden University Evolutionary and AI •not fully up to date, but clearly shows relations •boundaries much more fuzzy nowadays 15
  • 16. Discover the world at Leiden University CI and games: a field forming • 1999: Blondie24, learning checkers with CI • 2005 first Computational Intelligence in Games (CIG) conference • IEEE TCIAIG Journal (Transactions on CI and Artificial Intelligence in Games) since 2009 • EvoGames track in Evo*, DETA track in GECCO since 2009/2011 • many “neighbor” conferences, etc. AIIDE, FDG, gameai/nucl.ai conf. (not strictly CI, but CI welcome) • general approach target oriented, not technique oriented • 2012: first Dagstuhl seminar on AI and CI in Games • 2015: second Dagstuhl seminar (integration) • 2017: third Dagstuhl seminar (AI-supported design) • 2018/2019: journal and conference renamed to IEEE Transactions on Games and IEEE Conference on Games • 2019: first Shonan seminar on AI for Games 16
  • 17. Discover the world at Leiden University one small step for [a] man…. what recently happened: 17 AlphaGo: Go (MCTS, expert knowledge, self-play) AlphaGoZero: Go (MCTS-, self-play) Deep Reinforcement Learning: ALE (pixel input, self-play) AlphaStar: StarCraft (MCTS?, pixel input, self-play, expert knowledge) AlphaZero: Go, Chess, Shogi (MCTS-, self-play) Go-Explore: Montezuma (population based exploration) Obstacle Tower Challenge (population based training) ForTheWin: Capture the Flag (2 speed rnns, population based learning, population based training)
  • 18. Discover the world at Leiden University combining EC and Deep Reinforcement Learning •some (limited) successes in weight optimization with EC •population-based learning: - GANs, multiple agent architectures - resembles co-evolutionary approaches, same problems (how to control) •population-based based training: - generating test cases as needed - procedural content generation (PCG), mainly done via EC methods •not really exploited yet: multi-modal/multi-objective approaches •maybe more things to transfer? 18
  • 19. Discover the world at Leiden University what I did in Game AI ●procedural generation of strategy game maps ● organizing StarCraft I competitions, survey ● some works on balancing (also using EC): ○there are usually many solutions in balancing ○the problem is very multi-modal ●recently: facet orchestration (how to combine several generators, e.g. visuals, audio) ●very recently: transfer AlphaAnything to new problems (not from the games domain) ●understand how it works on small games 19
  • 20. Discover the world at Leiden University Obstacle Tower Challenge current DRL challenge for artificial general intelligence (AGI) • levels increase in difficulty • levels procedurally generated • decorations change • includes minigames (Sokoban) runs in 2 phases: - qualification (gone) - finals (up to July 15th, 2019) https://www.aicrowd.com/challenges/unity- obstacle-tower-challenge 20
  • 21. Discover the world at Leiden University bla Procedural Content Generation
  • 22. Discover the world at Leiden University search-based vs. generate-and-test 22
  • 23. Discover the world at Leiden University Evolutionary Algorithms • initial solutions needed: existing or randomly generated solutions • fitness function shall measure the targeted properties (designer input) • matching variation operators need to be established 23
  • 24. Discover the world at Leiden University why using EAs in PCG? pros: - flexible, only needs good utility function - not very sensitive concerning noise/mistakes - anytime method: we can obtain a result whenever needed cons: - optimality cannot be guaranteed (except for very few special cases) - stagnation phases of unpredictable length (break?) - depends on matching representation and operators - slow on simple problems 24
  • 25. Discover the world at Leiden University back in 2009: galactic arms race 25
  • 26. Discover the world at Leiden University how are the new weapons produced? • very small population, 3 solutions (candidates, individuals) • interactive selection: player removes existing solution when new one is generated • weapons are created via a parametrized particle system by complex neural network • cgNEAT: neural network topology optimized by means of evolutionary algorithm 26
  • 27. Discover the world at Leiden University a sample gallery 27
  • 28. Discover the world at Leiden University StarCraft I map evolution (2009-2012) 28
  • 29. Discover the world at Leiden University RTS map design with Sentient Sketchbook Preuss, Liapis, Togelius. Searching for Good and Diverse Game Levels. IEEE CIG 2014. 29
  • 30. Discover the world at Leiden University gamesbyangelina and the procjam • Michael Cook’s AI that is designed to intelligently design video games: http://www.gamesbyangelina.org/ • PROCJAM: yearly game jam with strong emphasis on PCG http://itch.io/jam/procjam • coming soon: facet orchestration 30
  • 31. Discover the world at Leiden University Project seminar nd balancing
  • 32. Discover the world at Leiden University Game AI hot topics • procedural content generation (PCG) - large game worlds are expensive to make - represent specific (designer) styles • believability of non-player character (NPC) behavior - more humanlike behavior (also evoke and show emotions) - better cooperation of units (team AI) - human/AI interaction • personalization - preference modeling (what do they like?), player type analysis - dynamic adaptation of game content and mechanics balancing is everywhere 32
  • 33. Discover the world at Leiden University balancing project with BlueByte • prototype from BlueByte • ZombieVillage game, related to tower defense • several parameters: can we balance automatically? 33
  • 34. Discover the world at Leiden University balancing project with BlueByte • number of (most important) parameters reduced to 5 • interchange between manual and automated (integrated) balancing • core result: both steps needed, many different solutions exist Marlene Beyer, Aleksandr Agureikin, Alexander Anokhin, Christoph Laenger, Felix Nolte, Jonas Winterberg, Marcel Renka, Martin Rieger, Nicolas Pflanzl, Mike Preuss, Vanessa Volz: An integrated process for game balancing. CIG 2016 34
  • 35. Discover the world at Leiden University tackling a complex RTS Mike Preuss, Thomas Pfeiffer, Nicolas Pflanzl, Vanessa Volz: Integrated Balancing of an RTS Game: Case Study and Toolbox Refinement. CIG 2018 35
  • 36. Discover the world at Leiden University bla nd general game AI/competitions
  • 37. Discover the world at Leiden University can we also have general (game) AI? • instead of this: • we rather want this: game 1 game 2 game 3AI 3 AI 2 AI 1 develop for develop for develop for game 1 game 2 game 3 AI 1 learn to play learn to play learn to play 37
  • 38. Discover the world at Leiden University GVGAI ENVIRONMENT • general video-game playing (GVGP) succeeds general game-playing • targeted at more complex, non-deterministic games • language first shaped during Dagstuhl seminar CI and AI in games 2012 • we want to obtain controllers that can play “any game” • important publications: John Levine, Clare B. Congdon, Michal Bída, Marc Ebner, Graham Kendall, Simon Lucas, Risto Miikkulainen, Tom Schaul, and Tommy Thompson. General Video Game Playing. Dagstuhl Follow-up, 6:1–7, 2013. Tom Schaul. A Video Game Description Language for Model-based or Interactive Learning. In Proceedings of the IEEE Conference on Computational Intelligence in Games, Niagara Falls, 2013. IEEE Press http://www.gvgai.net 38
  • 39. Discover the world at Leiden University GVGAI competition • build general AI controllers • learn to play many games, not one • many different challenges • many successful controllers use Monte-Carlo Tree search • but Rolling Horizon EA also an option • but how can this work? • “traditional track”: 1-player planning • employs forward model (action simulation, may be non-deterministic) • new: 2-player planning, learning (no forward model) • now also: level generation, rule generation 39
  • 40. Discover the world at Leiden University some example games Infection: Missile Command: Pacman 40
  • 41. Discover the world at Leiden University VGDL – game and level definitions 41
  • 42. Discover the world at Leiden University A SAMPLE (RANDOM) CONTROLLER 42
  • 43. Discover the world at Leiden University Rolling Horizon EA for search based planning • basic idea: evolve whole sequence of moves at once, apply only first action • hybrids with MCTS like playouts are also possible • simple extension with shift buffer (only remove latest action) performs better Raluca Gaina, Simon Lucas, Diego Perez. Rolling Horizon Evolution Enhancements in General Video Game Playing. IEEE CIG 2017 43
  • 44. Discover the world at Leiden University current competitions held at COG 2019 • 4th Angry Birds Level Generation Competition • Bot Bowl I • Fighting Game AI Competition • First TextWorld Problems: A Reinforcement and Language Learning Challenge • Geometry Friends Game AI Competition • General Video Game AI Competitions (Learning Track) • Hanabi Competition • Hearthstone AI competition • MicroRTS AI Competition • Short Video Competition • StarCraft AI Competition • Strategy Card Game AI Competition lots of demo and explanation videos included: http://ieee-cog.org/competitions_conference/ 44
  • 45. Discover the world at Leiden University conclusions •EC is part of AI and has always been •Procedural Content Generation getting very important for games industry but also for Deep Learning •evolutionary approaches have great potential for improving other algorithms, e.g. RHEA (rolling horizon EA) instead of MCTS •3 important applications of EC algorithms in game AI: PCG, balancing, search-based planning 45
  • 46. Discover the world at Leiden UniversityDiscover the world at Leiden University Don’t panic! 46