Anticipation and Emotion: a low-level approach to Believability Carlos Martinho, João Gonçalves, Ana Paiva Instituto Superior Técnico MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Outline IST MindRACES Scenario Aibo in the domotic household Low-level Anticipatory Affective Architecture Test Environment: Aini Integrates Anticipation and Attention Integrates Anticipation and Emotion Integrates Uncertainty as Meta-Anticipation Preliminary Evaluation of the Architecture Future Work in MindRACES MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 IST Scenario Scenario 1 - Task 3: “Looking for an object”: Takes place in a household environment where Aibo, the synthetic dog, “lives”. Several distractors, will be added to difficult the task and provide with opportunities for Aibo to “play in character” and be evaluated in terms of believability. MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 IST Scenario Context Our Competences: Provide with an anticipatory affective component, a low-level approach that provides with believable behaviour while an agent is searching for an object. Simulation only. Need components for real-world integration. Missing Links: Learn to find an object in an occluded environment (IDSIA fovea + UW-COGSI learning strategies). High level anticipatory affective cognitive reasoning component (ISTC-CNR BDI extension with expectation) for dealing with unplausibility, curiosity, cautiousness... Using reasoning by analogy to complement the search process (NBU). MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Virtual AIBO Toolbox Main purpose is to assess if the physical restrictions of Sony AIBO are adequate for the expression of believable behaviour. MindRACES, First Review Meeting, Lund, 11/01/2006
AIBO Domotic Environment AIBO sleeps near the children while they play in the living room when it senses the arrival of their father from work. Aibo will run to the front door and starts barking, anticipating the arrival of the owner. MindRACES, First Review Meeting, Lund, 11/01/2006
AIBO Domotic Environment AIBO will monitorize the domotic system and will give clues on what relevant events are occuring inside the household. MindRACES, First Review Meeting, Lund, 11/01/2006
Low-Level Architecture (D5.1) Sensor Effector Emotivector (lower-cognition anticipatory affect) SENSATIONS Agent Processing (BDI extension) (higher-cognition anticipatory affect) EMOTIONS MindRACES, First Review Meeting, Lund, 11/01/2006
AINI Environment Testbed anticipatory believability MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 The Word Puzzle Game MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Anticipation ? 0.4 0.3 0.2 time MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Anticipation Expected Value 0.5 0.4 0.3 0.2 time MindRACES, First Review Meeting, Lund, 11/01/2006
Anticipation and Attention Posner 1980 Müller and Rabbit 1989 Expected Value 0.5 0.4 0.3 Sensed Value 0.2 0.2 time SURPRISE = automatic reaction to a mismatch (Castelfranchi 2005) MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Attention in Action [demo] MindRACES, First Review Meeting, Lund, 11/01/2006
Anticipation and Emotion ? 0.4 0.3 0.2 time MindRACES, First Review Meeting, Lund, 11/01/2006
Anticipation and Emotion Some signals may have a search value MindRACES, First Review Meeting, Lund, 11/01/2006
Anticipation and Emotion search 1.0 current distance ? 0.4 0.3 0.2 time MindRACES, First Review Meeting, Lund, 11/01/2006
Anticipation and Emotion search 1.0 expected distance current distance 0.5 expected reward 0.4 0.3 0.2 time MindRACES, First Review Meeting, Lund, 11/01/2006
Anticipation and Emotion search 1.0 sensed distance sensed punishment 0.4 0.3 0.2 0.2 time MindRACES, First Review Meeting, Lund, 11/01/2006
Anticipation and Emotion Expectated Qualia vs Sensed Surprise (S) Positive Increase (S+) Positive Reduction ($+) Negative Increase (S-) Negative Reduction ($-) MindRACES, First Review Meeting, Lund, 11/01/2006
Anticipation and Emotion Harlow and Stagner (1933) Emotion versus sensation Young (1961) Emotion as process in hedonistic continuum Hammond (1970) Existence / absence of stimuli Millenson (1967) Intensity versus name Example of Sensation: $+ Harlow and Stagner - discontentment Hammond - Distress Millenson - negative unconditioned stimulus MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Emotions in Action MindRACES, First Review Meeting, Lund, 11/01/2006
Uncertainty and Salience Management Emotivector Salience Management Strategies: Winner takes-all: idle and restrictive Resource ordering: wasted in low relevance Treshold limit: which value to use? Meta-Anticipation MindRACES, First Review Meeting, Lund, 11/01/2006
Uncertainty and Salience Management Model( M ) Error Prediction = Uncertainty Model( S ) System S Environment E Meta-anticipatory System MindRACES, First Review Meeting, Lund, 11/01/2006
Uncertainty and Salience Management predicted error 0.4 0.3 0.2 time MindRACES, First Review Meeting, Lund, 11/01/2006
Uncertainty and Salience Management non-relevant signal 0.4 0.3 relevant signal 0.2 time Introduces uncertainty as error-prediction (resilient to white noise - Schimidhuber) Extension to 9 sensations (using neutral-based sensations) MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Prediction Comparative Evaluation of different algorithms: Polynomial Extrapolation (cubic curves) Error-Based learning Kalman filtering Statistical Limitation PID based prediction MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Prediction Winner Kalman-filtering simplification + 2-phase recirculation algorithm + statistical limitation MindRACES, First Review Meeting, Lund, 11/01/2006
Evaluation: Word Puzzle Scenario MindRACES, First Review Meeting, Lund, 11/01/2006
Preliminary Evaluation Believability-oriented evaluation Pre-evaluation with 10 representants of 5 user-groups: from 5 to 79 years-old of both sexes different familiarities with computer systems Divided the experiment in 2 fases: Training: navigation, attention, emotion Word Puzzle: 2 control, 1 emotivector-based, 1 common sense algorithm used in games All 10 subjects succeeded at the task but (surprisingly) only with emotivector-based approach! MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Future Work Evaluation of AINI scenario foir believability with users Assert relevance for the community Implementation of architecture in AIBO Evaluation of domotic scenario for believability with users Integration of high-level Anticipatory Affect (ISTC-CNR) Integration of real-world search functionality: IDSIA (fovea), UW-COGSI (search) and NBU (analogy) Evaluation of the integration (real time constraints) MindRACES, First Review Meeting, Lund, 11/01/2006
MindRACES, First Review Meeting, Lund, 11/01/2006 Questions? MindRACES, First Review Meeting, Lund, 11/01/2006