Opponent modeling

1
Opponent modeling

• Use of opponent modeling
• Opponent modeling techniques

Rik Boss Maarten Schadd
2
Use of opponent modelling

• To improve strength of play
• (beating the human player)
• But also
• Assisting the human player
• Training the human player
• Providing a proper level of opposition
• However is it needed?

3
Use of opponent modelling

• Complex commercial computer games
• Black White
• Neverwinter Nights
• But also
• Chess
• Poker
• Other board games

4
Beating the human player

• Anticipating opponent moves (chess)
• Enabling realistic game play (poker)
• Preventing repetitive behavior (first person
  shooters)
• Extensive opponent modeling in complex
  commercial games remains difficult.

5
Assisting the human player

• Buddy behavior
• Taking over player characters
• Providing realistic
• NPCs

6
Assisting the human player

• Buddy behavior
• In games like
• Brothers in Arms
• Proper behavior
• No micro-control
• More complex actions
• Massive online games (world of warcraft)
• Providing offline character continuation
• Providing multiple player entities

7
Training the human player

• Teach a game to a human
• Give feedback during gameplay
• Train skills and strategy
• Offering new challenges

8
Training the human player

• Examples
• Tutorials
• Personalized training missions in RTS
• Board game
• teaching programs
• Sparring partners
• (Tekken)

9
No more need?

• Games are improved all the time
• Standard AI becomes better and better
• Games are played online
• Scripting is available
• However
• Boring NPCs will always be available
• Opponent modeling offers personalization.

10
Human Intentions

• Idle
• Recover
• Snipe
• Engage Static
• Pullback
• Engage Forward
• Move up
• Sneak
• Ambush

11
Observations

• 21.0 Idle Static, Stand Rifle
• 21.2 Idle Static, Stand Rifle
• 21.4 Sneak Static, Crouch, Rifle, Threat,
  TLooksAwayP
• 21.6 EngageForward ForwardFast, Rifle, Aim

12
Adapting chances

• Naive Bayes
• Training
• Classification

13
Player Model

• Remebering what player does.
• Selected situations
• Success of effect
• NewValue aOldValue(1-a)Observed
• Query player model

14
N-Gram Statistical Model

• ABABABABB
• 75 chance of A after a B
• Bigram
• TriGram
• UniGram

15
N-Gram Statistical Model

• Predict players behaviour
• Choose best counter action
• Combining different N-Grams

16
GoCap

• Action Database
• Set of conditions
• Condition (currenthealth/maxHealth)
• Training

17
GoCap

• Bins
• Count observations
• Predicting

18
Reputation System

• NPC behavior
• Worldwide Reputation
• Local Reputation
• Forgetting

19
Reputation

• Event Announcer
• Intensity
• Unknown Data
• Conversation

20
Preference Model

• Preference Function
• V(s) 5players health -7number of
  enemies
• Inititialize weights
• Update weights
• Comparing game states

21
Preference Model

• PM-search
• Alpha Beta Tree search

22
Probabilistic Model

• Player classes
• Offensive Defensive
• Assign Chances
• Easy to adapt
• Robust
• Alpha Beta Tree search

23
Conclusions

• Many different means of opponent modeling are
  available
• Opponent modeling is widely applicable to games

24
References

• Articles published in AI Game Programming Wisdom
  (2002-2006), Charles River Media, Hingham, MA
• Sterren (van der), W. (2006). Being a better
  buddy Interpreting the players Behavior, pp
  479-494.
• Houlette Stottler, R. (2004). Player Modeling for
  Adaptive Games, pp 557-566.
• Laramee, F. D. (2002). Using N-Gram Statistical
  Models to Predict Player Behavior, pp 596-601.
• Alexander, T. (2004). GoCap Game Observation
  Capture, pp 579-585.
• Alt, G., King, K. (2004). A Dynamic Reputation
  System Based on Event Knowledge, pp 427-435.
• Other sources
• Herik (van den), H.J., Donker H.H.L.M., Spronck
  P.H.M., Opponent Modelling and Commercial Games.