Character AI: Sensing & Perception

Transcription

1 Lecture 21 Character AI:

2 Take Away for Today Sensing as primary bottleneck Why is sensing so problematic? What types of things can we do to improve it? Optimized sense computation Can we improve sense computation performance? Can we share sensing between NPCs? Sense event matching What are events and how are y represented? What is advantage of an event system? 2

3 Review: Sense-Think-Act Sense: Perceive world Reading game state Example: enemy near? Think: Choose an action Often merged with sense Example: fight or flee Act: Update state Simple and fast Example: reduce health? 3

4 Recall: Sensing Performance Sensing may be slow! Consider all objects Example: morale n knights, n skeletons Knights fear skeletons Proportional to # seen Count skeletons in view O(n) to count skeletons O(n 2 ) for all units Time per tick 3 units 2 units 1 unit 4

5 Recall: Sensing Performance Sensing may be slow! Consider all objects Example: morale n knights, n skeletons Knights fear skeletons Proportional to # seen Count skeletons in view O(n) to count skeletons O(n 2 ) for all units Time per tick 3 units 2 units 1 unit 5

6 Aggregation Idea taken from databases Unordered set of information Combine into single value Used in statistical analysis Examples: sum, avg, mode Decomposable Aggregates Split set up into subsets Aggregate on each subset Combine values from subsets Only for some aggregates

7 Aggregation Idea taken from databases Unordered set of information Combine into single value Used in statistical analysis Examples: sum, avg, mode Decomposable Aggregates Split set up into subsets Aggregate on each subset Combine values from subsets Only for some aggregates avg = 8 (16,2) avg = 6 (12,2) avg = 8 (16,2) avg = 6 (12,2) 7

8 Aggregation Idea taken from databases Unordered set of information Combine into single value Used in statistical analysis Examples: sum, avg, mode avg = 8 (16,2) avg = 6 (12,2) avg = 7 (28,4) Decomposable Aggregates Split set up into subsets Aggregate on each subset Combine values from subsets Only for some aggregates avg = 8 (16,2) avg = 6 (12,2) avg = 7 (28,4) 8

9 Aggregation Idea taken from databases Unordered set of information Combine into single value Used in statistical analysis Examples: sum, avg, mode Decomposable Aggregates Split set up into subsets Aggregate on each subset Combine values from subsets Only for some aggregates avg = 8 (16,2) avg = 6 (12,2) avg = 8 (16,2) avg = 6 (12,2) avg = 7 (28,4) avg = 7 (28,4) avg = 7 (56,8) 9

10 Aggregation Idea taken from databases Unordered set of information Combine into single value Used in statistical analysis Examples: sum, avg, mode Decomposable Aggregates Split set up into subsets Aggregate on each subset Combine values from subsets Only for some aggregates avg = 8 (16,2) avg = 6 (12,2) avg = 8 (16,2) avg = 6 (12,2) avg = 7 (28,4) avg = 7 (28,4) avg = 7 (56,8) 10

11 AI and Aggregation Trees Slide courtesy of Dave Mark Number of Allies Strength of Allies Number of Enemies Strength of Enemies Proximity to Base My Health Proximity to Leader 11

12 AI and Aggregation Trees Slide courtesy of Dave Mark Number of Allies Strength of Allies Number of Enemies Strength of Enemies Allied Strength Enemy Strength Proximity to Base My Health Proximity to Leader 12

13 AI and Aggregation Trees Slide courtesy of Dave Mark Number of Allies Strength of Allies Number of Enemies Strength of Enemies Allied Strength Threat Ratio Enemy Strength Proximity to Base My Health Proximity to Leader 13

14 AI and Aggregation Trees Slide courtesy of Dave Mark Number of Allies Strength of Allies Number of Enemies Strength of Enemies Allied Strength Threat Ratio Enemy Strength Proximity to Base My Health Proximity to Leader Urgency 14

15 AI and Aggregation Trees Slide courtesy of Dave Mark Number of Allies Strength of Allies Number of Enemies Strength of Enemies Allied Strength Threat Ratio Enemy Strength Proximity to Base My Health Proximity to Leader Urgency My Morale 15

16 AI and Aggregation Trees Slide courtesy of Dave Mark Number of Allies Strength of Allies Number of Enemies Strength of Enemies Allied Strength Threat Ratio Enemy Strength Proximity to Base My Health Proximity to Leader Urgency My Morale Retreat % 16

17 AI and Aggregation Trees Slide courtesy of Dave Mark Number of Allies Strength of Allies Number of Enemies Strength of Enemies Allied Strength Enemy Strength My Health Proximity to Leader Threat Ratio Computable independent Urgency Proximity to Base My Morale of NPC Retreat % 17

18 Related Approach: Tactical Managers Invisible NPC Assigned to NPC Group Performs all thinking Tactical Manager NPCs just follow orders Applications Protecting special units Flanking NPC NPC NPC NPC Covering fire Leapfrogging advance 18 Thinking and Acting

19 Protecting Special Units Slide courtesy of Dave Mark 19 Thinking and Acting

20 Protecting Special Units Slide courtesy of Dave Mark 20 Thinking and Acting

21 Protecting Special Units Slide courtesy of Dave Mark 21 Thinking and Acting

22 Protecting Special Units Slide courtesy of Dave Mark 22 Thinking and Acting

23 Protecting Special Units Slide courtesy of Dave Mark 23 Thinking and Acting

24 Protecting Special Units Slide courtesy of Dave Mark 24 Thinking and Acting

25 Protecting Special Units Slide courtesy of Dave Mark 25 Thinking and Acting

26 Protecting Special Units Slide courtesy of Dave Mark 26 Thinking and Acting

27 Protecting Special Units Slide courtesy of Dave Mark 27 Thinking and Acting

28 Performance: Loop Inversion Normal Sensing Inverted Sensing sound1 NPC Loop Sense Loop sound2 sight1 sight1 Loop over all NPCs Check what NPC senses Loop over sensations Send se to each NPC 28

29 Performance: Loop Inversion Normal Sensing Inverted Sensing sound1 NPC Loop Sense Loop sound2 sight1 sight1 Loop over all NPCs Check what NPC senses Loop over sensations Send se to each NPC 29

30 Sense Events Event: encoded sense data Tagged with sense type Pre-aggregated Information self-contained O(n) data is aggregated O(1) to combine w/ NPC Sensing is event matching Each event has a type NPCs register for a type Send NPC registered events Check if event is relevant 30

31 Sense Event Matching sound sight sound Register events of interest Event Handler Game Loop sound smell 31

32 Sense Event Matching Notify of any matching events Event Handler Game Loop Check for any matching events 32

33 Event Handling in LibGDX MessageDispatcher Send with dispatchmessage delay (0 if immediate) sender (can be null) target (null for subscribers) type (user defined int code) data (object, like Box2D) Subscribe with addlistener NPC to receive message Type (int) to subscribe to Telegram Stores event message Entries of dispatchmessage Except for delay value Preaggregated sense in data Received by Telegraph Interface for receiver Implemented by NPC One method: handlemessage 33

34 Event Handling in LibGDX MessageDispatcher Send with dispatchmessage delay (0 if immediate) sender (can be null) target (null for subscribers) type (user defined int code) data (object, like Box2D) Subscribe with addlistener NPC to receive message Type (int) to subscribe to Telegram Stores event message Entries of dispatchmessage Except for delay value Preaggregated sense in data Received by Telegraph Interface for receiver Implemented by NPC One method: handlemessage 34

35 Recall: S-T-A Architecture Actor1 Controller Actor2 Controller Compute Sensing GameState Actor1 Actor2 35 Thinking and Acting

36 Recall: S-T-A Architecture Actor1 Controller Actor2 Controller Compute Sensing GameState Actor1 Actor2 36 Thinking and Acting

37 Recall: S-T-A Architecture Actor1 Controller Actor2 Controller Compute Thinking GameState Actor1 Actor2 37 Thinking and Acting

38 Sensing: Perception Groups Vision: limited field of view Gives exact object location, information Limited by obstacles and range Little information (motion) at periphery Sound: omni-directional Gives direction & distances Requires you track sounds actions make Smell: omni-directional No direction or distance; proximity only Requires you track smells actions make 38

39 Case Study: Thief Series 39

40 40

41 Line-of-Sight in Thief Long Distance Focused View Peripheral Vision Short Distance 41

42 Line-of-Sight in Thief Long Distance Focused View Motion Detection Peripheral Vision Short Distance 42

43 Sounds in Thief Easier than vision Primarily distance-based Decays probabilistically Tag with level of interest Sounds can be blocked Not same as line-of-sight Use alternate level map Or tag your visible map Not physically realistic Echoes? Reflections 43

44 Sounds in Thief Easier than vision Primarily distance-based Decays probabilistically Tag with level of interest Sounds can be blocked Not same as line-of-sight Use alternate level map Or tag your visible map Not physically realistic Echoes? Reflections 44

45 Sounds in Thief Sounds are general purpose Resuable framework Code is lightweight Encodes or senses Example: Smell Treated as pseudo-sound Generate like any sound * sniff * Again, ignores or factors * sniff * Wind direction Masking smells 45

46 Custom Data in Events Lightweight Heavyweight Memory Target Memory Target ThiefExposed Reference to target ThiefExposed Copy of target 46

47 Custom Data in Events Lightweight Advantages Fast to create event No additional memory Disadvantages Must be used immediately Lost over frame boundary Ideal for fast decisions Heavyweight Advantages Can persist past frame Can retain as memory Disadvantages Must allocate memory Object ownership is tricky Ideal for communication 47

48 Communicating Senses 48

49 Communicating Senses First Hand LOS Sight & Sound 49

50 Communicating Senses First Hand LOS Sight & Sound First Hand LOS Sight & Sound 50

51 Communicating Senses First Hand LOS Sight & Sound First Hand LOS Sight & Sound Second Hand Sight & Sound 51

52 Communicating Senses First Hand LOS Sight & Sound First Hand LOS Sight & Sound Second Hand Sight & Sound? 52

53 Alertness: Active Senses High Alert First Hand LOS Sight & Sound First Hand LOS Sight & Sound Second Hand Sight & Sound 53

54 Alertness: Active Senses High Alert First Hand LOS Sight & Sound Medium Alert First Hand LOS Sight & Sound Second Hand Sight & Sound 54

55 Alertness: Active Senses High Alert First Hand LOS Sight & Sound Medium Alert First Hand LOS Sight & Sound Second Hand Sight & Sound High Alert First Hand Sound 55

56 Alertness: Active Senses High Alert First Hand LOS Sight & Sound Medium Alert First Hand LOS Sight & Sound Second Hand Sight & Sound High Alert 56

57 Alertness: Active Senses High Alert First Hand LOS Sight & Sound Medium Alert First Hand LOS Sight & Sound Second Hand Sight & Sound High Alert First Hand Sound 57

58 Spatial Optimizations Restrict to nearby NPCs Have detection range Limits events sensed Easy to combine with event matching system Works in both directions Nimbus: can see radius Aura: can be seen radius Area of interest management 58

59 Spatial Optimizations Restrict to nearby NPCs Have detection range Limits events sensed Easy to combine with event matching system Works in both directions Nimbus: can see radius Aura: can be seen radius Area of interest management 59

60 Spatial Optimizations Restrict to nearby NPCs Have detection range Limits events sensed Easy to combine with event matching system Works in both directions Nimbus: can see radius Aura: can be seen radius Area of interest management 60

61 Thief: Sense Events and Aggregation Position Lighting Movement Exposure Sound System Visibility Sound Queue Viewcone Selector Viewcone Non-specific Spatial Events Look Listen Sense Pulse Receiver Game Mechanics and Configuration Inter-Agent Communication Ramp Up Delay Cool-down Capacitor Inter-Agent Observation 61 Sense Links

62 Summary Sensing is most expensive part of AI Each character looks at every object in game Often leads to O(n 2 ) behavior (bad!) Can optimize sense garing Aggregation is amenable to parallelization Can piggyback some data onto pathfinding Event matching inverts sensing problem Creation of sense makes a data event Forward event to relevant NPCs 62