probabilistic smart terrain dr. john r. sullins youngstown state university

39
Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

Upload: mavis-owens

Post on 17-Jan-2018

221 views

Category:

Documents


0 download

DESCRIPTION

John Sullins Youngstown State University Probabilistic Smart Terrain ICTAI Smart Terrain Solves complex navigation problems in real time

TRANSCRIPT

Page 1: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

Probabilistic Smart Terrain

Dr. John R. SullinsYoungstown State University

Page 2: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

2

Outline

• What is Smart Terrain?

• Why do we need to add probabilities?

• Estimating expected distances to objects that meet character needs

• Plausibility benchmarks and experimental results

• Adding learned knowledge during exploration

• Hierarchical application to games

Page 3: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

3

Smart Terrain

• Solves complex navigation problems in real time

Page 4: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

4

Smart Terrain

• Characters have “needs”– Example: hunger

• Objects in world meet needs– Example: refrigerator with food inside

• Characters move towards objects that meet needs

Page 5: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

5

Smart Terrain

• Objects meets needs transmits “signal”– Signal weakens with distance– Signal moves around objects

Page 6: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

6

Smart Terrain

• Characters follow signal to objects– Move in direction of

increasing signal

– Only need to compute map once when level created

– No need for complex navigation!

Page 7: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

7

Need for Probabilities

• Smart terrain can result in implausible actions

– Room character has never visited– Contains empty refrigerator

• Does not transmit signal• Character ignores it

– Not plausible behavior!

Page 8: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

8

Probabilistic Smart Terrain

• Objects broadcast signal of form“I meet need n”

“I may meet need n with probability P ”

• Probability = uncertainty that object meets need

• Character might explore uncertain objects along path

Page 9: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

9

Probabilistic Smart Terrain

• Theoretical goal:Move to closest object with highest probability

• Problem: Optimizing two separate criteria!

• Actual Goal: Plausible behavior for characters

Meets “hunger”

need with P = 0.7

At distance 8

Meets “hunger” need with P = 0.6

At distance 6

Page 10: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

10

Expected Distances

• Expected number of tiles character must travel to reach object that fulfills need

• Use to determine which tile to move to next– Compute expected distance for four surrounding tiles– Move to surrounding tile with lowest value for

expected number of tiles to travel

Page 11: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

11

Expected Distances

P(t): probability no objects within t tiles meet need

P(t) = (1 – pi ) (Equation 1)

where di < t

• Based on:– di: distances to each object i– pi: probabilities each object i meets need– Assumption of conditional independence

Page 12: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

12

Expected Distances

• t < 6: P(t) = 1• 6 ≤ t < 8: P(t) = (1 – 0.6) = 0.4• t ≥ 8: P(t) = (1 – 0.6)(1 – 0.7) = 0.12

Distance: 6 Prob: 0.6

Distance: 8Prob: 0.7

Page 13: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

13

Expected Distances

Expected distance from tile T to tile that meets need

E(T) = Σ P(t) (Equation 2) t

t < 6: P(t) = 16 ≤ t < 8: P(t) = 0.4

t ≥ 8: P(t) =0.12

Page 14: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

14

Expected Distances

• Problem: Sum could be infinite• Solution: Limit t to some tmax tmax > di i

tmax

E(T) = Σ P(t) (Equation 3)

t

Page 15: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

15

Expected Distances

• Compute expected distance E(T) for all tiles T• Character moves to adjacent tile with lowest E(T)

Page 16: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

16

Plausibility Benchmarks

• Goal for games:Non-player characters should behave plausibly– Move in direction that “makes sense” to player

• Benchmarks for plausible behavior:– Objects similar in either distance or probability – Group of objects in same direction– Objects that meet need with complete certainty

Page 17: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

17

Plausibility Benchmarks• Objects at same distance move to higher probability

• Objects with same probability move to closer one

Page 18: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

18

Plausibility Benchmarks• Nearly same distance move to much higher probability

• Nearly same probability move to much closer object

Page 19: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

19

Plausibility Benchmarks

Aggregate probabilities benchmark:• Multiple objects > single object with higher probability

– Assumption of conditional independence

Page 20: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

20

Plausibility Benchmarks

Complete Certainty benchmark:• Single object with probability = 1 >

multiple objects with probability < 1

Page 21: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

21

Learned Knowledge

• Probabilities changed when object reached– Object meets need probability becomes 1– Does not meet need probability becomes 0

• Should affect future actions

Refrigerator empty

Move towards another goal

Page 22: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

22

Learned Knowledge• Changing global map affects all characters

– Will also appear to have learned this knowledge

New character enters room

Also ignores empty refrigerator

Page 23: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

23

Learned Knowledge

• Each character stores own world model– Belief object meets needs– Initially based on probabilities– Modified when objects explored

Refrigerator R1 70%Refrigerator R2 80%

Object Belief object meets need

0%

Page 24: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

24

Learned Knowledge

• Each object propagates raw data to tiles– Probability it meets need– Distance to that tile

Page 25: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

25

Learned Knowledge• Character examines surrounding tiles

– Modify probabilities using world model

– Compute expected distances for each

Page 26: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

26

Hierarchical Smart Terrain

• More realistic scenario:– Know whether objects meet needs– Don’t know if object is present in given area

• Go to entrance of most likely area• If object present, move to it.• Otherwise, move to another area

Page 27: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

27

Hierarchical Smart Terrain

• “Area attractors” at entrances to rooms– Broadcast to entire level– Probability object that meets need is in room– Probability set to 0 when reached by character

• Objects in room– Signal range = size of room– Probability = 1 if present in room

Page 28: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

28

Hierarchical Smart Terrain

• Compute expected distances from area attractors• Move to “best” room

Page 29: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

29

Hierarchical Smart Terrain

• Object is present in area:– Now in range of object, probability meets need = 1– Character will move directly to object

Page 30: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

30

Hierarchical Smart Terrain

• Object is not present in area:– Set probability of area attractor = 0– Character will move to next plausible attractor

Page 31: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

31

Conclusions

• Probabilities added to Smart Terrain algorithm

• Characters move to adjacent tile with shortest expected distance to a tile that meets need

• Algorithm produces plausible behavior for benchmarks

• Probabilities overridden by learned knowledge

• Hierarchical algorithm for realistic play

Page 32: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

32

Ongoing Work

• Algorithm modification to avoid local minima

• Characters with multiple needs at different levels– Low-probability object that meets critical need– High-probability object that meets less critical need– Which to move towards?

• Objects that change over time– Empty refrigerator now may be restocked in future

Page 33: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

33

Local Minima

Caused when paths to low probability objects overlap

Overlap in paths to P=0.23 objects

Tiles nearer to object appear farther away

Page 34: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

34

Local Minima

Solution: Weight estimated tiles by distance tmax

E(T) = Σ P(t) k t t

• Nearby objects “appear” even closer

• Any weight k > 0 seemsto work

Page 35: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

35

Multiple Needs

• Characters can have multiple needs– Hunger, Fun

• Some needs more critical than others– Hunger = 10 Fun = 5

• Objects may only partially fulfill needs– Donuts: Hunger-7– Cookies: Hunger-3– TV: Fun-6

• Needs increase over time (each tile traversed)– Hunger += 0.5 per tile Fun += 0.2 per tile

Page 36: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

36

Multiple Needs

• Goal: Minimize total “discontentment”

Σ (needj)2

j

• Problem: Balancing different factors

Page 37: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

37

Multiple Needs• Terminology:

– n j = current level of need j

– di j = distance to object i

– pi j = probability object i meets need j

– ai j = amount that need j decreased by if it meets need)

– c j = increase in need j for each tile traversed

• Expected decrease in need j caused by all objects i = Σ pi

j ai

j within t tiles i

Page 38: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

38

Multiple Needs

• Expected level of need j if move t tiles:

max (0, n j + tc j

- Σ pi j ai

j ) where di < t

i

• Expected discontentment if move t tiles:

Σ (max (0, n j + tc j

- Σ pi j ai

j ))2 where di < t j i

Page 39: Probabilistic Smart Terrain Dr. John R. Sullins Youngstown State University

John Sullins Youngstown State University

Probabilistic Smart Terrain ICTAI 2009

39

Multiple Needs

• Total expected discontentment at given tile:

tmax

Σ Σ (max (0, n j + tc j

- Σ pi j ai

j ))2 t j i

• Compute for surrounding tiles• Move to tile with lowest expected discontentment