Machine Learning in Games George Dolbier CTO Interactive Media IBM
Machine Learning in GamesGeorge DolbierCTO Interactive Media IBM
What is Machine Learning?
Software thatLearns and Produces
WithoutExplicit
Programming
ASK DISCOVEREXPLORE DECIDE VISUALIZE
Ask questions for greater insight
Natural language dialogue & robotics
Image Video Audio spam OCR
Evidence-based decisions with with traceability
Consolidate and visualize
Loci of Machine Learning
Try this for fun Google “Personality Insights Demo” or use this QR code to go tohttps://personality-insights-livedemo.mybluemix.net/
4
Machine Learning in Games
Creating better play experienceCase Study
Balance and Engagement
One of the hardest problems in game design is Play Balance.
Poor progression ramps, too steep, too shallow, awkward jumps, sited as #1 factor in dis-engagement
Questions posed to the team:
• Can an ML be used to balance the play experience?
• How can you tailor an experience for a specific player?
• Can the individual’s telemetry data be used to make better balance decisions?
• Can Curated content be delivered to players programmatically to produce a better overall experience?
Case Study : Plight of the ZombieSituation:
• Simple puzzle game• Dozens of short 30 to 90 second experiences• Designed by level designers• Each level given a “curated/seeded” difficulty
For this initial use case 3 variables are considered:1. Time it took player to complete level2. Number of retries the player took to complete level3. Curated difficulty metric assigned to each level
John O’Neil: Sparkplug Games
How tradeoff is used to balance the gameShow Watson Tradeoff call
ResultPlayer plays through tutorial levels, and seeded “easy” levels
3 Tutorial levels, 2 “easy” levels
By the 5 level enough data has been collected for Watson to begin suggesting ramp in difficulty
Next StepsFor this simple use case next obvious steps will be:Use all player telemetry to correct difficulty settingsHave Watson identify play patterns in player baseHave Watson tell developers progression, engagement, trends
− How long to players play− Are players playing longer, shorter− What is triggering in game purchase or conversation
Have Watson automatically generate freebies or offers based on struggle
10
Machine Learning in Games
Truly Interactive NPCCase Study
Real time interaction with a fictional character
A Long time desire in the industry
Can you create a character, that a human can interact with?
Can you have a conversation with a character that includesA BackstoryAn AttitudeLikes, DislikesConversation Not just scripted QnA tree
https://www.thesuspect.com/
The Suspect - Immersive chat thriller
Second screen app with synchronised news alerts
and live 3D brain scan
Main screen with dynamic video, AI chat powered by Watson, gamified experience, and transmedia storyline
Gamified conversation with simulated points, level and rank
Contextually-served video to match suspect’s responses
Case Study : The Suspect
Can we “Throw everything we can” into a characterAnd through natural conversation, bring the player into the character’s world
3 Types of information make up the personality1.Mind Map2.Traditional Q and A (Word) 3.Conversation loops (how does character react to repeated
questions)8 to 10 people total, core team maxed at 6, calendar time 18 months
Core team focused development 4 months
Guy Gadney: Lead Developer for “The Suspect”
Use of Conversational technologyConversational chat bot associated with Brazilian TV showAverage session was 20 minutes8% of chats lasted for over an hour (Target Audience)Site traffic increased 15%. This lead to an increase in advertising revenue around the project's pages.
But Wait! There is More:Alpha Go! https://deepmind.com/alpha-goProject Aries http://goo.gl/eMAQMuGuy Gandy HowWeGetToNext article on chatbots https://goo.gl/6xTCafMedical Minecraft http://goo.gl/dD8BMxFashion Design http://goo.gl/Ps9EBCGoogle machine learning recipes : https://goo.gl/9k2ASxMari/o Using evolution to train neural networks https://goo.gl/Jxf73VArtomatix ML for texturing https://artomatix.com/
Thank [email protected]@NoirTalonhttps://www.linkedin.com/in/georgedolbier