Designing Human-like Video Game Synthetic Characters through Machine Consciousness

Transcription

1 Designing Human-like Video Game Synthetic Characters through Machine Consciousness Raúl Arrabales, Agapito Ledezma and Araceli Sanchis Computer Science Department Carlos III University of Madrid Toward a Science of Consciousness 2009 PolyU, Hong Kong, 12th June 2009

2 Contents Introduction Objectives Research Environment description Hypothesis and Ongoing Work Open problems Conclusions Questions 2

3 Introduction Context: Machine Consciousness Research. Experimentation domain: First Person Shooter (FPS) video games. Working hypothesis: To improve Human-like synthetic character design using cognitive theories of consciousness. 3

4 A bit of History From robotics to computer games: 4

5 Context (UTIII and UT2004) Unreal Tournament Epic Games, Inc. 5

6 Environment (I) Bot Behavior Controller Sensory Data Processing Sensory Inputs Body State Basic Behaviors Movement Module Animation Module Atomic Actions Virtual World 6

7 Environment (II) CERA UT2004 Viewer UT2004 Server GameBots 2004 CERA-CRANIUM CCR / DSS Runtime.NET Framework Translated Pogamut 2 UT2004 Client (Human player) UT2004 Client (Spectator) 7

8 Technologies Unreal Engine. Unreal Script. GameBots. Pogamut 2. Robotics Developer Studio. CERA-CRANIUM. 8

9 Main Objectives To understand how cognitive skills associated with consciousness can be combined. To understand what does human-like means. 9

10 Other objectives To test Machine Consciousness models. To explore the role of qualia in behavior generation. To develop practical measures of consciousness based on behavior. To develop more engaging synthetic characters. 10

11 Video games as research tools Benefits: Practical experimentation environment. Computationally affordable and rich. Easy assessment of global agent behavior. Focus on high-level control. Human (virtual) interaction. 11

12 Video games as research tools Drawbacks: Lack of real physical world interaction. Results cannot be directly applied to robotics or other domains. Noise and uncertainty are generally ignored. Prone to misinterpretations. 12

13 Research Approach Bots controlled by flexible cognitive architecture to explore the integration of cognitive skills. But, How? Using consciousness as inspiration. 13

14 Cognitive Skills (0) First step: Identifying cognitive skills associated with consciousness. Is this a reductionist approach? Does this mean that phenomenal aspects are ignored? 14

15 Cognitive Skills (I) Executive Function High level control for maintaining and implementing a goal. Attention. Set Shifting. Planning. Coordination. 15

16 Cognitive Skills (II) Emotions Emotions as low level reactions. Feelings as representation of the former. Conscious feelings as second order representations of the relation between the self and feelings. 16

17 Cognitive Skills (III) Theory of Mind (Lewis, 2003) I know (proprioceptive sensing). I know I know (self-consciousness). I know you know (inter-subjectivity). I know you know I know (development of Machieavelian strategies). 17

18 How to integrate everything? Using a cognitive architecture But, how the different skills can be combined effectively? Let s use consciousness as the source of design inspiration. 18

19 Computational Model Contexts Focus of Attention Working Memory Specialized Processors Interim Coalition 19

20 Cognitive Architecture CERA-CRANIUM CERA Virtual Bot World Physical Layer Mission-specific Layer Core Layer Sensor Services Motor Services Sensors Actuators Accessible environment Reachable environment 20

21 CERA-CRANIUM (I) Modulation Commands Workspace Workspace Core Layer Single Percepts Complex Percepts Mission Percepts Physical Layer Mission-Specific Layer 21

22 CERA-CRANIUM (II) Core Layer Goals and meta-goals (keep positive emotional state). Long term memory. Learning mechanisms. Model of self. Models of others. 22

23 What should be done first? What makes the agent behave humanlike? ConsScale proposes a developmental path. 23

24 Small Demo (video) 24

25 Open Problems Accurately testing for consciousness. Lack of accurate verbal report. Practical definition of human-like behavior. Judging protocol. Other problem domains. 25

26 Conclusions (I) When the judge is just a passive spectator bots are more likely to appear as humans. When the judge participates in the game, interaction is better assessed, and bots look less smart. Some human players don t pass the Turing test. 26

27 Conclusions (II) Human-like versus Conscious-like. To improve believability (it is always a ToM game ). Very good performing bots are unlikely to be considered human. Poorly performing bots are unlikely to be considered human. 27

28 Conclusions (III) Current state of the art Multiplayer online games are more appealing because opponents are humanlike. Current Synthetic characters are not human-like (see 2k BotPrize 2008 contest). 28

29 Conclusions (IV) Considering fixed repertories of behaviors or isolated cognitive skills doesn t seem to be a good approach. We argue that using the consciouslike approach is helpful in the quest for human-like characters. 29

30 Future Work Turing Test for Bots 2K BotPrize Competition IEEE Symposium on Computational Intelligence. September. Milan. Italy. Challenge: narrative bot. 30

31 Related talks CERA-CRANIUM: A Test Bed for Machine Consciousness Research. International Workshop on Machine Consciousness. Establishing a Roadmap and Metrics for Conscious Machines Development. 8 th IEEE International Conference on Cognitive Informatics. 31

32 Thank you 32