Interaction of Simulated Actors with the Environment. By Henk W. M. Gazendam Groningen University Twente University OS8, June 23/24, 2005

Transcription

1 Interaction of Simulated Actors with the Environment By Henk W. M. Gazendam Groningen University Twente University OS8, June 23/24, 2005

2 Overview The ACT-RBot project ACT-R Exploring ACT-R improvement Encoding the environment as affordances Emotion and awareness 2

3 The ACT-RBot project Aim: Multi-actor simulation based on ACT-R Team: Martin Helmhout, Gijs Roest, Henk Gazendam, René Jorna ACT-R rebuilt in Java Client-server system Each actor = client Simulated environment = server 3

4 The ACT-RBot project Actor communication based on TCP/IP sockets FIPA-ACL XML Flexible memory organization Perception 2D movement 4

5 The ACT-RBot project Data collection in database Experiment: actor interaction in 2D world shows formation of social constructs To explore: improvement of interaction of ACT-RBot with environment & other actors 5

6 ACT-R The ACT-R cognitive architecture Goal-oriented production system Declarative memory & procedural memory Perception & action buffers Conflict resolution at subsymbolic level Production with the highest expected utility fires Boltzmann factor Learning: Subsymbolic: Continuously varying activation of declarative chunks and productions; forgetting possible Symbolic: Encoding of perception data, storing popped goals, production compilation 6

7 ACT-R 5.0 Intentional Module (not identified) Goal Buffer (DLPFC) Declarative Module (Temporal/Hippocampus) Retrieval Buffer (VLPFC) Productions (Basal Ganglia) Matching (Striatum) Selection (Pallidum) Execution (Thalamus) Visual Buffer (Parietal) Visual Module (Occipital/etc) Manual Buffer (Motor) Manual Module (Motor/Cerebellum) Environment

8 ACT-R Problems of ACT-R: Physical grounding The apparent complexity of our behaviour is largely a reflection of the complexity of the environment (Simon, 1969) A system that is intelligent needs its representations grounded in the physical world (Brooks, 1991) The world is its own best model The trick is to sense it often Cognition can use the semiotic Umwelt instead of internal representations Cognition needs to recognize affordances for adequate behaviour 8

9 ACT-R Problems of ACT-R: Symbol grounding Symbol manipulation should be about something and the symbols should acquire their meaning from reality (Vogt, 2002) Searle s Chinese Room Argument The use of Peircean triads can overcome the symbol grounding problem The actor needs to learn the link between symbol, object and interpretant (habit) 9

10 ACT-R Problems of ACT-R: Social behaviour Time scale ACT-R s basic time unit is 50 milliseconds ACT-R operates in the cognitive band and lower rational band (100 milliseconds 2 min) We are interested in the social band (17 min 12 days) Components ACT-R has declarative chunks, productions and goals as atomic components No knowledge about larger (molecular) structures like social norms 10

11 Exploring ACT-R improvement Organizational semiotics Encoding affordances seems basic step Affordances (Gibson, 1979) are properties of the environment that make possible or inhibit animal activity People develop repertoires of behavior tuned to affordances (Stamper, 2001) Perception of affordances is a process of construction of Peircean triads A semi-indexical sign An object or pattern in the environment A behaviour pattern (interpretant) 11

12 Exploring ACT-R improvement Simulation of emotion Breazeal (2004) has designed a robot motivated by emotional drives: Need to interact with people (social drive) Need to be stimulated (stimulation drive) Need to occasionally rest (fatigue drive) The emotion system Adds affective information to incoming perceptual information Maintains an affective state, Determines emotive response 12

13 Exploring ACT-R improvement How to improve the ACT-RBot actor: Encode signals from the environment as affordances Elaborate to social constructs Emotion subsystem affords goal switching Active environment Sends signals, reacts on actions Enforces laws of nature Takes care of time synchronization 13

14 Encoding the environment as affordances Elaboration Sign Encoding Possible Actions Perception Goal Handling Signal Acting (Object) 14

15 Encoding the environment as affordances Affordance encoding Object sends characteristics (XML) Actor encodes received sign as declarative chunk Registers time and place Assigns affordance identity Registers characteristics Determines affordance class Connects to possible actions 15

16 Encoding the environment as affordances Affordance = Object + Possible Action + Sign Object : Action : Sign 1:1:1 basic affordance 1:n:1 basic affordance n:1:1 situation n:m:1 situation with script Situations and scripts must be seen as affordances Situation encoding = affordance + Roles Scripts Script encoding = affordance + Actions as {phase, role, action} 16

17 Encoding the environment as affordances Actions enabled by affordances Actor needs basic knowledge of possible actions afforded by an affordance class Actions taken are dependent on social context & applicable norms Based on experience, the knowledge of suitable actions is refined 17

18 Encoding the environment as affordances Action category Move body Manipulate object Social action Actions Move Rest Take, Put, Make Eat Give, Receive Move, Dance, Work together Speak, Listen Write, Read, Send, Receive1 Affordance classes Body +Terrain Body Thing, Food Food Actor + Thing, Food, Document Actor + Body + Terrain Actor + Message Document + Actor 18

19 Encoding the environment as affordances Elaboration to social constructs Actor determines social context and norms based on Encoded situations Other actor behaviour encoded as scripts Social context elements Community Situation type Own role & task Script Phase in script 19

20 Encoding the environment as affordances Norm elements Name Type Social Context Condition Deontic operator Action 20

21 Emotion and awareness The basic loop: Perception Encoding & elaboration (new: affordance encoding; social construct elaboration) Emotion (new: intentional module) Awareness (new: intentional module) Goal handling Action 21

22 Emotion and awareness Emotion Awareness Encoding Goal Handling Perception Action (Environment) 22

23 Emotion and awareness The emotion subsystem: Encodes body signals Monitors incoming information and adds affective information to it Maintains an emotional state based on this affective information Encodes emotional state as emotion affordances that can be picked up by the awareness subsystem 23

24 Emotion and awareness Drive Safety Fatigue Social Stimulation Actions Move (avoid collision) Move (towards something) Eat, Rest Any other action Social action in accordance with norm Social action not in accordance with norm, non-social action Do task, Explore, Evaluate & learn Any other action State change Towards safe Towards fearful Towards energetic Towards tired Towards content Towards unhappy Towards excited Towards bored 24

25 Emotion and awareness Awareness The awareness buffer contains information about the current context and emotional state It is used for selection of the norms that influence behaviour The awareness handler Triggers elaboration and integration of new perceptual information Updates the awareness buffer Activates goal change based on the emotional state 25

26 Emotion and awareness Behaviour according to social norms Is encouraged by the social drive Doing tasks, exploration, evaluation and subsequent learning Are encouraged by the stimulation drive 26

27 Emotion and awareness Learning Exploration Action + Encoding of affordances Imitation Encoding of interaction patterns of other actors as situations and scripts + Action Knowledge transfer Listening to messages and reading documents + Action Evaluation Identifying and naming generalized action patterns and affordance subclasses + Action 27

28 Conclusion What can ACT-R learn from organizational semiotics? Affordance (Peircean triad) encoding Context awareness Behaviour encouraged by social norms What can organizational semiotics learn from ACT-R? Affordances extended to situations Social norms dependent on social context Social context consisting of community, situation, role, task, script, and phase 28

29 Encoding the environment as affordances Affordance (sign) types Physical affordance Basic affordance Physical situation Document Emotion affordance Social construct Situation Community Role Script Phase in script Norm 29

30 Emotion and awareness Emotions and goal switching Overall goal Improve the emotional state At least to acceptable If perfect, explore & learn If emotional state is not acceptable, goal stack is emptied until the overall goal gets focus Popped goals are remembered 30

31 Emotion and awareness Steps to generality Affordance encoding Elaboration to social constructs Integration into declarative memory Check whether affordance is already known Identifying and naming generalized action patterns and affordance subclasses 31