Building a Sports AI Architecture

1
Building a Sports AI Architecture

• Ushhan D. Gundevia
• November 8, 2004

2
Terry Wellmann

• He has been programming since 1983
• Responsible for architecting and writing AI for
  Microsofts NBA Inside Drive franchise
• He was also the lead programmer on All-Star
  Baseball for N64
• Has a Computer Science Degree from Purdue
• terry.wellmann_at_high-voltage.com

3
If you would like to know more

• http//xbox.ign.com/articles/372/372418p1.html
• http//www.mobygames.com/developer/sheet/view/by_g
  enre/developerId,126714/

4
Challenges

• Real teams spend countless hours practicing
  together Why?
• To improve an individuals skills and abilities
• To train a group of independently thinking
  individuals to instinctively react to a situation
  in a manner predictable by other teammates
• To stimulate cohesive group decision making in a
  game is difficult

5
Things to keep in mind

• Keep it simple
• Break down decisions into various levels
• Spend time on planning
• Dont be afraid to make mistakes
• Do not underestimate the power of Randomness

6
Agent Plans

• Examples of Offensive, Defensive and Shared plans
• Pass
• Shoot
• Drive to basket
• Run the play
• Rescue a trapped ball handler

7
Pseudo code for Agent Plan

• Class AgentPlan
• float EvaluateInitiation()
• float EvaluateContinuation()
• void Initiate()
• void Update()

8
Why float?

• The EvaluateInitiation and Evaluate-Continuation
  functions return float values
• Always building complex systems, with several
  plans compared to determine which one is used
• Each plan is a building block in a complex system
• Each plan individually evaluates the current
  situation to determine its feasibility
• Returns a value between -1.0 and 1.0
• Execute plans if value gt 0
• To strongly encourage a plan, return a value gt 1

9
Team Management

• Its implemented using a FSM
• Each state encapsulates a different goal and a
  set of responsibilities
• Basketball can be assumed to have 3 kinds of
  states
• Offensive States
• Defensive States
• Common States

10
Offensive and Defensive States

• They mirror each other
• Responsible for all aspects of the game play that
  the user can interact with.
• They are ?
• Inbound
• Transition
• Frontcourt
• Rebound
• Recover Loose Ball
• Free Throw

11
Common States

• Neutral situations when the ball is not in play
  and neither team is in offense or defense
• They are ?
• Pre-game
• Tip-off
• Time-out
• Quarter Break
• Substitution
• Post Game

12
State Transition and Flows
13
Pseudo Code - Frontcourt Offense State

• Class FrontCourtOffense public TeamStateBase
• protected
• enum TBallHandlerPlans
• eBallHandlerShot, eBallHandlerPass,
• eBallHandlerDrive, eBallHandlerPlayMovement,
• enum TNonBallHandlerPlans
• eNonBallHandlerRescueBallHandler,
• eNonBallHandlerGetOpenForPass,
• eNonBallHandlerSetPick
• eNonBallHandlerPlayMovement
• public

14
Whats that?

• The key elements are the two enumerated types
  that define priority order
• We might also have 2 different update functions
  for a BallHandler and a NonBallHandler for
  simplicity
• In some situations priorities might not be
  enough, me may need some explicit if-then logic
• Any Idea where?

15
Pseudo Code for Update Fn

• for (each plan to evaluate)
• //get the proper evaluation value
• if (planToCheck currentPlan)
• evalValue EvaluateContinuation()
• else
• evalValue EvaluateInitiation()
• //is the current plan the best option
• //simple priority order plan selection
• if (evalValue gt bestEvalValue)
• bestEvalValue evalValue
• bestPlan currentPlan
• //handle situations the priority method cant
• else if (bestValue gt 0.0f bestPlan shot)
• //is the drive plan a valid option
• if (currentPlan drive evalValue gt 0.0f)
• //set the best plan to drive, but leave the
  bestValue as it is Why?

16
Agent AI

• A well thought-out collections of Utility
  Functions and data used by the plans to evaluate
  and execute the things they represent

17
Shot Plan

• Taking a closer look at the shot plan
• Three things to keep in mind
• Potential success rate
• Players skills and abilities denoted by rating
  values
• Penalties e.g. defensive pressure applied to
  the player
• Type of shot
• Is Dunk possible?
• How far is he from the basket?
• Does he have a clear path to the basket?
• Can he dunk?
• Is Lay-up possible?
• Same questions as above, but every professional
  NBA player can lay-up
• Last option is Jump shot
• Tell the player how to perform that shot

18
High-Level Agent AI

• Command issuing functions
• Decisions that require more logic then a simple
  random no.
• Whether to do a flashy 360 dunk or a simple one
• Wrapper for Agent Mechanics

19
Agent Mechanics

• Low-Level AI
• Manage and select the animations
• Simple decisions that require no more than Random
  nos.
• Like commands given to people who have no
  knowledge about basketball, but can follow simple
  directions. e.g. walk, jog, run

20
Conclusion

• Sports games present a unique set of challenges
• The rules are well defined and allow little room
  for creativity
• Professional athletes have personalities and
  styles that are recognized by millions of fans
• User should be able to capture the abilities,
  personalities and emotions of the athletes being
  stimulated

21
References

• Building a Sports AI Architecture AI Game
  Programming Wisdom 2
• Learning Goals in Sports Games Jack van
  Rijswijck
• www.gdconf.com/archives/2003/Van_Ryswyck_Jack.doc
  
• A very nice paper dealing with learning in Sports
  Games, using Soccer as the base.

22

• Thank You