50
Deep Learning for Game Development Magnus Nordin Technical Director, EA SEED

Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Embed Size (px)

Citation preview

Page 1: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Deep Learning

for

Game Development

Magnus NordinTechnical Director, EA SEED

Page 2: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 3: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 4: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Future Worlds

Deep LearningVirtual Humans

Prototypes

Page 5: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

f( ) = cat

Page 6: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

f( ) =

f( ) =

”A person flying a kite

on a beach”

”A coffee, please.”

Page 7: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

f( ) =”A coffee, please.”

f( ) =

Page 8: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

f( ) =This bird is red and brown in

color, with a stubby beak

f( ) =This flower is pink, white,

and yellow in color, and

has petals that are striped

StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks, Zhang et al, 2016

Page 9: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

f( ) =

Page 10: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Reinforcement Learning for

Game AI

Page 11: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

1987 2015

Page 12: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Reinforcement Learning

Observations/Rewards

AgentEnvironment

Goal

Actions

Learning by doing

Page 13: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Bad Balls

Our hero

Eat! +1 point

Avoid! -5 points

Page 14: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 15: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 16: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 17: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

EA Trailer

Page 18: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 19: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 20: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 21: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Single-Action vs Multi-Action

Page 22: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Objective area

The agent

Health

The opposition

Supplies

Page 23: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

50M steps of training

Page 24: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 25: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

100M steps of training

Page 26: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 27: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 28: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Generalization

Page 29: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 30: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Challenges

Observation (vision vs state)

Combining with classical game AI

Giving designers controlGoals

Imitation

Execution (GPUs are typically busy with graphics ☺)

Page 31: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Machine Learning Gaming Use Cases

Page 32: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 33: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Physics

Physics Forests: Real-time Fluid Simulation using Machine Learning, Ladicky et al., 2015, www.physicsforests.com

Page 34: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

GAN geometry

Interactive Example Based Terrain Authoring with Conditional Adversarial Networks, Guérin et al, 2017

Page 35: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Representation Learning and Adversarial Generation of 3D Point Clouds, Achlioptas et al., 2017

Page 36: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

• Needs to be Full of Life (Spore connection?)

Page 37: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Emergence of Locomotion Behaviours in Rich Environments, Heess et al., 2017

Page 38: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

El último vals

A Neural Parametric Singing Synthesizer, Blaauw & Bonada, 2017

Page 39: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Audio-Driven Facial Animation by Joint End-to-End Learning of Pose and Emotion, Karras et al., 2017, NVIDIA

Page 40: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

The Future

Page 41: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 42: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Geek & Sundry, D&Diesel

Page 43: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

[clip]

• Live role playing

Page 44: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

MMORPG

Page 45: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

MMO(RPG)s

• Not much RPG [pic of WoW raid]

• hence ”MMOs”

Page 46: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

True Role Playing

Page 47: Magnus Nordin at AI Frontiers: Deep Learning for Game Development
Page 48: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

f( ) =

f( ) =

Page 49: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Sopranos © HBO 2006

Page 50: Magnus Nordin at AI Frontiers: Deep Learning for Game Development

Games will change more in the next 5 years than they have in the last 45 years

Magnus NordinTechnical Director, EA SEED