Mo#va#on and Affect in MicroPsi

Transcription

1 New Des3na3ons in Ar3ficial Intelligence Mo#va#on and Affect in MicroPsi MAS S66 New Des+na+ons in Ar+ficial Intelligence Goals and Direc+ons for Future Research

2 Components for Cogni.ve AI Memory Resolution Perception Action selection Action Securing rate Arousal Motivational system: Motive Selection and Decision making Selection threshold Urges/drives

3 Layers of Cogni.on Reflec3ve Delibera3ve Reac3ve

4 Columns of Cogni.on Percep3on Cogni3ve Processing Ac3on

5 Cogni.ve Grid Reflexive Percep3on Delibera3ve Percep3on Reac3ve Percep3on Meta- Management Planning, Reasoning Reflexes Management Ac3on Delibera3ve Ac3on Reflexive Ac3on

6 Conceptual Analysis: HCogAff (Sloman 2001) 12/2/15 FutureAI

7 Cogni.ve Ar.ficial Intelligence Methods should focus on components and performances necessary for intelligence: Universal Representa#ons: Grounded neuro-symbolic representa3ons (integrate both symbolic and distributed aspects) (Semi-) Universal Problem Solving: Learning, Planning, Reasoning, Analogies, Ac3on Control, Reflec3on... Universal Mo#va#on: Polythema3c, adap3ve goal iden3fica3on Emo#on and affect Whole, testable architectures

8 Modeling Mo.va.on in a Cogni.ve Architecture General intelligence needs General Mo.va.on Mo3va3onal system structures cogni3on Mo3va3onal dynamics: physiological, social and cogni3ve drives Inten3on selec3on and ac3on control Mo3va3on vs. affect

9 MicroPsi architecture PSI theory Principles of Synthetic Intelligence (Dörner 1999; Bach 2003, 2009)

10 Acknowledgements Work on MicroPsi2 is collabora3ve effort: - Ronnie Vuine, Dominik Welland, Priska Herger, Jonas Kemper are contributors to the current version - Architecture/concepts have been inspired by Dietrich Dörner, Aaron Sloman, Marvin Minsky, Stan Franklin and many others - Support from Humboldt University of Berlin, University of Osnabrück (Ins3tute for Cogni3ve Science), Berlin School of Mind and Brain, Harvard Program of Evolu3onary Dynamics, MIT Media Lab

11 MicroPsi Principles Neuro-Symbolic architecture Agents implemented as spreading ac3va3on networks Unified representa3ons, different sets of opera3ons All representa3ons are grounded Meaning is abached to representa3ons by mo3va3on

12 Implementa.on: MicroPsi 2 (Bach, Welland, Vuine, Herger 12, 14) 12/2/15 FutureAI

13 Goals in MicroPsi Goal: situa3on or ac3on that afforts to sa3sfy a need Aversive goal: situa3on or ac3on that frustrate a need All behavior is directed on sa3sfying an appe33ve goal or avoiding an aversive goal Needs are predefined, goals are learned

14 From Needs to Behavior Priming + Modulation Memory Needs Urge Signals Learning Perception Decision Making Action

15 Mo.va.on in MicroPsi Pleasure and distress: Change of a demand is reflected in pleasure or distress signal Strength is propor.onal to amount of change Pleasure and distress signals deliver reinforcement values for behavioral procedures and episodic sequences and define appe++ve and aversive goals.

16 Mo.va.onal System drive = demand + urge indicator target current level Water s water urge indicator

17 Mo.va.onal Learning mo3ve = urge + goal situa3on target s water current level Water urge indicator goal

18 Mo.va.onal Learning mo3ve = urge + goal situa3on target s water + current level Water urge indicator goal situa3on aversive situa3on

19 Physiological needs Thirst Hunger Rest Warmth Libido à Survival as emergent property

20 Social needs Affilia3on (Aben3on from others, external legi3macy) Internal legi3macy Nurturing (caring for others) Affec3on Dominance

21 Cogni.ve needs Competence: Skill acquisi3on (epistemic competence) Coping/control ability (general competence) Effect genera3on Uncertainty reduc3on: Explora3on Aesthe3cs: S3mulus oriented Structure oriented (abstract aesthe3cs)

22 Needs and urges

23 Mo.va.onal Learning associa3on by learning: urge indicator demand s i w 1 + goal situa3on change indicator Δs i V+ A+ V- A- w 2 aversive situa3on valence associator

24 Mo.va.onal Learning retrogradient reinforcement protocol chain goal demand urge indicator s i w 1 + change indicator Δs i V+ A+ valence associator

25 Mo.va.onal Learning Mo3vator: situa3ons leading up to goal = plan goal urge s i w 1 + autonomous regula3on V+ A+ Δs i valence associator

26 Mo.va.onal Learning Inten3on: goal 1 urge goal 2 s i autonomous regula3on goal 3... goal n

27 Mo.va.onal learning

28 Mo.ve selec.on

29 Need parameters Strength: rela3ve importance Decay: rate of replenishment Gain: effect of sa3sfac3on Loss: effect of frustra3on different configura3on of need parameters = different personality traits

30 Modula.on in PSI/MicroPsi

31 Primary modulators

32 Compare: Affec.ve dimensions (Wundt 1910)

33 ARen.onal modulators

34 Modulator dynamics

35 Modulator parameters Baseline Range Vola3lity Dura3on Different modulator parameter configura3ons = different temperaments

36 Emo.ons as directed affect + Modula.on Examples: Fear: an3cipa3on of aversive events (à neg. valence) + arousal Anxiety: uncertainty (à neg. valence) + low competence + arousal, high securing behavior (frequent background checks)

37 Emo.ons as directed affect + Modula.on Examples: Anger: Perceived obstacle (usually agent) manifestly prevented reaching of an ac3ve, mo3va3onally relevant goal (àneg. valence), sanc3oning behavior tendency (à goal relevance is re-directed to sanc3oning of obstacle), arousal, low resolu3on level, high ac3on readyness, high selec3on threshold Sadness: Manifest preven3on from all conceived ways of reaching ac3ve, relevant goal, without relevant obstacle (à neg. valence), support-seeking behavior (by increased demand for affilia3on), low arousal, inhibi3on of ac3ve goal à decreased ac3on readyness

38 Emo.ons as directed affect + Modula.on Examples: Pride: high competence (à low securing rate), high internal legi3macy, likely coincidence with high external legi3macy Joy: high arousal + high perceived reward signal from sa3sfying a demand Bliss: low arousal + high perceived reward signal from sa3sfying a demand (since physiological demands ojen involve high arousal, mostly related to cogni3ve demands, such as aesthe3cs)

39 Individual Varia.ons by Parameterizing Possible grounding of personality proper3es (FFM): Openness: apprecia3on of art and new ideas, curiousity Conscien#ousness: rulefollowing vs. chao3c Extraversion: tendency to seek s3mula3on by environment and others Agreeableness: tendency for coopera3veness and compassion Neuro#cism: emo3onal stability, effect of failure to self-confidence

40 Needs and Big Five

41 Player personality types Richard Bartle (1996): Hearts, Clubs, Diamonds, Spades: Players Who suit MUDs

42 Mo.va.on and personality Personality proper3es can be modeled as mo3va3onal variability Affilia3on Uncertainty reduc3on Competence

43 Needs and player types (with S. Tekovsky)

44 Mo.va.on in MicroPsi All ac3ons are directed on goals or avoidance of aversive goals All goals are established through learning how to sa3sfy needs All decisions are based on strengths of urges and chances to sa3sfy corresponding needs Personality differences are the result of parametriza3on of the mo3va3onal system

45 Emo.on in MicroPsi Affec3ve states are configura3ons of cogni3on, by modulators Primary modulators: arousal, valence, agression Aben3onal modulators: focus, securing rate, resolu3on level High-level emo3ons are determined by an affec3ve state that is directed on mo3va3onal content