330 likes | 506 Views
The Effect of Trust and Team Personality On M ultiagent T eamwork Performance. Feyza Merve Hafızoğlu Boğaziçi University January 13, 2010. Outline. Motivation Composed Trust Teamwork Trust Tools ART Experimental Results - 1 Agents with Personality Performance Evaluation
E N D
The Effect of Trust and Team Personality On Multiagent Teamwork Performance Feyza Merve Hafızoğlu Boğaziçi University January 13, 2010
Outline • Motivation • Composed Trust • Teamwork Trust Tools • ART • Experimental Results - 1 • Agents with Personality • Performance Evaluation • Experimental Results -2 • Conclusion 2
Motivation • Most real life needs are satisfied by composite services. • Furniture workshop:a designer, a carpenter, an assembler, a painter. • Factors on performance: trust, personality, leadership • Consider teams with: • Varying levels of task interdependence • High role and task differentiation • Distributed expertise • Definitions: • Composed Service: Teamwork • Service: Task • Group of providers: Team 3
Composed Trust • Observation: Agent’s performance differs in teamwork environment and single service environment • Trust to a team rather than trust to an individual • A team of successful individuals SUCCESS ????? • How providers work well together? • Research Question: How much an agent can trust a team of agents? • Teamwork effect: Agent’s behavior depends on teamwork • Teamwork Trust Tools: Team Trust Model, Service Graphs 4
Modeling Experiences • Individual Trust Model: • How much can I trust Axfor Ty? • Represent expertise of individuals for tasks • Team Trust Model: • How much can I trust team(Ax, Ay, Az) for teamwork (Ta,Tb, Tc) ? • Agents are not penalized or rewarded individually, but as a team • Expertise: • A real number between 0 and 1 • Update according to performance of team/individual 5
Service Graphs (Yolum&Singh, 2006) • A service graph is a weighted, directed graph • Nodes for teamwork • The edges for transitions between teamwork • Teamwork, which have at least one common task, are connected to each other • The weights of the edges show how likely providers that are successful in a source node are likely to be successful in the target node 6
Team Formation Strategies 1stStrategy: TeamTrust Model TW5: T1, T2, T3 2ndStrategy: Service Graph, TeamTrust Model 2nd Strategy TW7: T2 TW2: T1, T3 3rdStrategy: Individual Trust Model 3rd Strategy 8 TW1: T1 TW7: T2 TW3: T3
Experimental Framework ART Testbed 9 How much can I trust agent X for era Y?
ART with Teamwork • Teamwork Structure: • Tasks (paintings) • # Tasks: 1-4 • Different task types (eras) in a teamwork • One agent is responsible for each task • Each task has a weight • True values are generated for each painting • Overall error: Weighted sum of individual errors • Weights table: Captures the characteristics for different teamwork • Only the simulator knows weights! 10
Example: Teamwork • A teamwork is created as the following: • Number of tasks: 3 • Task types (eras) : Era1, Era3, and Era7 • Weights: 0.45, 0.25, and 0.30 • True values: 1000, 2000, and 1500 • Appraised values: 1100, 1980, and 1560 • Simulator calculates the overall error: Error for the 1st task ………….(0.45 * (|1100 - 1000| /1000)) Error for the 2nd task…………..(0.25 * (|2000 - 1980| /2000)) Error for the 3rd task………...+(0.30 * (|1560 - 1500| /1500)) Error for teamwork……….….. Appraisar agents are informed about the overall error 0.0545 11
Noncooperative Behavior • Consider agents whose behavior may change based on the particular teamwork that it’s taking part in • Noncooperativeness list: Each agent has a finite list of teamwork in which it is going to be noncooperative • Noncooperativeness level: Shows the extent of cooperation, changing between 0 and 1 • Noncooperativeness in ART: Opinion provider agent sends the worst opinion creation order 12
Experimental Setup • TMA (Team Modeling Agents) uses Teamwork Modeling Tools • IMA (Individual Modeling Agents) uses Individual Trust Model • 3 TMAs, 3 IMAs, and 3 honest agents • Noncooperativeness Level: 0.3 • 200 timesteps • 100 repetitions • Compare TMA with IMAwrt: • Bank balance • Appraisal error 13
Result 1: Bank Balances • TMAs achieve higher bank balances than IMAs 14
Result 2: Appraisal Error • The appraisal error of TMAs is more stable than the appraisal error IMAs 15
Result 3: High Population • 10 TMAs, 10 IMAs, and 10 honest agents • Difference between bank balances increases • Appraisal error of TMAs is stable and lower than IMAs’error 16
Agents with Personality • Noncooperativeness doesn’t express real life situations successfully • Two important dynamics in teamwork: • Interdependency • Personality 17
Interdependency Relationships • Consider teamwork that has 3 tasks and exhibits one of five interdependency relationships 18
Big Five Personality Traits • Agreeableness (Likability) is tendency to be courteous,good-natured, cooperative, and trusting • Conscientiousness(Will to achieve) is tendency to beresponsible and hardworking, which are important characteristics for accomplishing work tasks. • Emotional Stabilityis tendency to be relaxed, secure,and calm. • Extraversionis tendency to be sociable, assertive, andactive. • Openness (Intellect and Intelligence)is tendency to beimaginative, cultured, and original. 19
Agents with Personality • Research Question: Can we charactarize successful teams wrt personality? • Four of Big Five Personality traits are applied in ART domain • Each trait is represented as a real number between 0 and 1. • The simulator determines trait levels randomly. • Agents are only aware of theirown personality trait level 20
Agreeableness in ART • Agreeableness measures the quality of interpersonal interaction. • Consider furniture workshop example: • Designer may be capricious • Designer may not trust to the carpenter when he objects • Interacting agents having high levels of agreeablenesswork well together • Multiply the performance of Aiby the average agreeableness level of Ai and Ai-1: 21
Conscientiousness in ART • Conscientiousness is tendency to being achievement oriented and persevering to accomplish tasks • Consider the furniture workshop example: • Each provider is responsible from his own task. • If the carpenter has a tendency to be undutiful, and careless, hisindividual performance would worsen. • Usually considered related to individual task performance • Opinion creation orderis affected by the conscientiousness level of Ai : 22
Emotional Stability, Extraversion in ART • Emotional stability is tendency to be secure and relaxed. • Agents having low emotional stability frequently exhibit unpredictable behaviors • Sending abnormal amounts (i.e.1) for opinion creationorder • Extraversion is tendency to being social and talkative. • Talkative appraiser agents send more certainty requests 23
Performance Evaluation • Teamwork (T1, T2, T3 ) is carried by the team (A1, A2, A3), respectively • Individual performances of A1, A2, and A3: 0.8, 0.7, and 0.5 • Agreeableness levels of A1, A2, and A3 : 0.8, 0.7, and 0.9 • Interdependency Factor • Overall performance:(0.40*0.80)+(0.35*0.42)+(0.25*0.22) = 0.52 • Agreeableness Factor 24
Experimental Setup • 50 homogeneous agents (IMA or TMA) • Agents have varying personalities • Levels of personality traits: low, medium, high • 300 timesteps • 50 repetitions • # tasks : 3 • Interdependency relationships: { 1, 2, 3, 4, 5 } • Teamwork instance: agent-task assignments, average team personality, performance, trust among team members • Focus on successful, unsuccessful, and most trusting teams 26
Result 1: Significance of Traits (IMA) • Higher conscientiousness and homogeneity of conscientiousness in successful teams • Trust among team members is not a guarantee of success 27
Result 2: Trust Model (TMA) • Results are parallel with the results of IMAs • Team trust model has an effect on trust among teams 28
Result 3: Occupational Groups • The significance of agreeableness increases as the interdependency level increasess 29
Result 4: Mature Teams • Emotional stability is significant for mature teams • Trust values of successful and unsuccessful teams are close to each other 30
Related Work • Barber studies a multi-dimensional trust-based mechanism for team formation. Candidate providers have different tendencies towards completing anassigned task. • Blizzard is an action-based approach for modelingthe environmentin ART. Blizzard outperforms agent-based approaches. • Aghaee studies the representation of fuzzy agents with dynamic personalities by using Big Five and extended 30 facets.A set of if-then rules are geared towardpersonality descriptors, factors, characteristics, and modifications. • Mount investigates the relationship between Big Five and three job performance criteria (job proficiency, training proficiency, andpersonnel data) for five occupational groups 31
Conclusion • Extend ART to handle teamwork • Develop a teamwork modeling strategy and implement it in an agent (TMA) • Show performance improvements using teamwork modeling • Model and implement Big Five Personality traits in an ART agent • Study emergent properties of these traits in ART teamwork • Implications: Teamwork trust model may be extended to model personality; e.g. Agents may model other agents’ conscientiousness and choose agents accordingly. 32
Publications • Feyza Merve Hafızoğlu and Pınar Yolum. Composing Trust: An Effective Trust Model for Multiagent Teamwork. In Proceedings of the AAMAS Workshop on Trust, 2009. • Feyza Merve Hafızoğlu and Pınar Yolum. The Effect of Big Five Personality Traits on Multiagent Teamwork.(in progress) 33