catholijn m. jonker
DESCRIPTION
Development and Application of Rich Cognitive Models and the Role of Agent-Based Simulation for Policy Making. Catholijn M. Jonker. BRIDGE : Development and Application of Rich Cognitive Models for Policy Making. Frank Dignum , Virginia Dignum , Catholijn M. Jonker. Policy. - PowerPoint PPT PresentationTRANSCRIPT
Development and Application of Rich Cognitive Models and the Role of Agent-
Based Simulation for Policy Making
Catholijn M. Jonker
BRIDGE: Development and Application of Rich Cognitive
Models for Policy Making
Frank Dignum, Virginia Dignum, Catholijn M. Jonker
Policy
• Policy introduction– Goal: noticeable change on the global level– Assumption: incentive for individuals to
change behaviour to intended new behaviour• Influencers of individual’s behaviour
– Dynamics of environment– Social circles (family, friends, work, culture …)– Personal circumstances
Example Policies• Anti-smoking ban:
– Aim: Healthy (work) environment– Result? Less bar revenues, civil disobedience
• VAT increases– Aim: More state revenues– Result? more black market, less revenues
• Higher demands on hospital hygiene– Aim: Better health– Result? superbugs
Levels of simulation / models• Macro-level to measure policy effect
– Model at macro level: • Averages over behaviour of individuals• Misses out on holistic effects
• Micro-level to allow variation in behaviours– Requires rich cognitive models
• Personality• Cultural differences
– Local variation• Personal circumstances• Social circles
Micro-macro simulation: zoom-in/zoom-out approach
The BRIDGE architecture
B
E
D
G
I
Inference method
personal orderingPreference
Cultural beliefs
Normative beliefs
Growth needs
deficiency needs
sense
act
generate
select plan
update
inte
rpre
t filter
plan select
direct
R
urges, stress
select
direct
over
rule
stimuli
explicit
implicit
BeliefsResponseIntentionsDesiresGoalsEgo
Support for Policy Makers
Old viewPolicy maker directly puts
policy at work in the society.
Agent-based simulation viewPolicy maker first tries out
the policy in the simulation
When would ABM help?• Agent should show
realistic human behaviour, with culture, social circles etc.
• If we can build agents that react realistically to any policy, then we solved the strong AI problem!
Agent-based simulation viewPolicy maker first tries out the
policy in the simulation
Policy – Effect examples• Goal: reduce garbage heaps• Policy: garbage bags are taxed• Effect: people dump garbage in nature
• Goal: Reduce “fat” from Ministry of Defense• Policy: Reduce budget• Effect: Minister announces Trade Fleet cannot
be protected from pirates
• Goal: Reduce risk of terrorist attacks• Policy: Forbid face covering clothing• Effect: Police officers refuse to enforce it
Our proposal• Identify stakeholders• Qualitative interviews with representatives of:
– target population– implementers of policy
Þ Possible implementations, possible reactions of targets, possible side effects
• Interview experts in psychology and national cultures to create XML file to link possible reactions to personality, culture, and circumstances
• Run simulations using XML file
Required Adaptations of Models• Additional info from
interviewed people – new actions and
decision rules– Adapt existing
decision rules when influenced by new actions
• Run simulation
policypossible reactions
possible side effects
Caveats
• Sensitivity analysis required of the – Basic agent model – Overall simulation model
• Validation!• Cannot predict, only explore possibilities
Theorizing
Theory,hypotheses
Gamesessions
Data,conclusions
Test design
Experimentalsetup
Gamingsimulation
Agentmodeling
Agent-BasedModel
Modelvalidation
Modelruns
Validationresults
Game design
Real world observations
Gaming simulation
Computer simulation
Theory
tests predictions based on
implements design of
implements mechanisms according to
validates mechanisms described by
tests predictions based on
Sensitivity Analysis of anAgent-Based Model of
Culture’s Consequences for Trade
Saskia Burgers, Gert Jan Hofstede, Catholijn Jonker, Tim Verwaart
September 9-10, 2010 - Treviso (Italy)
Sensitivity analysis
• Generally considered “good modeling practice”
• Actual parameter values are uncertain• A powerful tool in the process of model
verification and validation• Specific problems arise when performing
sensitivity analysis for agent-based models
Sensitivity analysis for ABM• Agent-based models may be very
sensitive to parameter changes in particular parts of parameter space:– Nothing may happen in large areas in the joint
parameter space– Areas may exist where the system responds
dramatically to slight changes• Parameters may significantly interact• Sensitivity may be studied for aggregated
individual level outputs
Influence of culture
• Culture modifies parameter values in the decision functions
• Describe culture based on Hofstede’s five dimensions of national cultures
• Relational attributes have different significance in different cultures:– Group distance– Status difference– Interpersonal trust
The role of parameters• Which areas in parameter space result in
realistic behavior?• In which areas of parameter space can
tipping points occur?• Which parameters have significant effects
for which outputs?• Which interactions between culture and
other parameters are important?• Are the answers different between
aggregate and individual level?
Results of sensitivity analysis (1/2)
• For many of the parameter sets drawn at random, no transactions occur
• No obvious regions in parameter space where transactions occur / no transactions occur
• Logistic regression: discover the parts of parameter space where transactions occur
• Zoom in on the regions in parameter space where interesting behaviour occurs
Results of sensitivity analysis (2/2)• Parameters that have significant effects can be
identified through meta-modeling, even for complex systems. However, the analysis is not straightforward.
• When keeping culture constant, straightforward methods for sensitivity analysis can be applied. Results differ considerably across cultures.
• Sensitivity of individual agents can differ considerably from aggregate level sensitivity.
Cross-validation of Multi-Agent Simulation withCultural Differentiation
Gert Jan Hofstede, Catholijn M. Jonker, Tim Verwaart
September 9-10, 2010 - Treviso (Italy)
Validation
• Why: to combat under-determinism• model M explains the behaviour of a
system S– Is M the only model to do so?
Cross-validation (Moss & Edmonds, 2005)
• Compare statistics of – Agent-based simulation– Simulated system at aggregate level
• Compare– Behaviour at individual level– Data from qualitative research
Human-like Agent behaviour
• Complexity requires compositionality• Process model composed of sub-process
models• Sub-models implement theories of
different aspects of behaviour:– Negotiation, trust, deceit …– Moods, emotions, affect, …
Culture complicates matters
• Social situations are culture-sensitive• Policies affect social situations• Policy making requires culture-sensitive
modelling
Our proposal to approach validation
• Complexity: Use compositionality– Validate sub-processes at lower compositional
levels• Qualitative Data: Use gaming simulations
– Played by humans for these sub-processes to gather data
Overall multi-agent
simulation
partialmulti-agent simulation
partialmicro
simulations
Compositional Cross-Validation
Example in Trade
• Trust & Tracing game to simulate trade chains
Producers Middlemen ConsumersRetailersProducers Middlemen ConsumersRetailers
Decision model within agent
determinetrade goal
selecttrade partner
negotiate
deliver
monitor and enforce
update beliefs
determinetrade goal
selecttrade partner
negotiate
deliver
monitor and enforce
update beliefs
Conclusion• BRIDGE: rich cognitive agents & support for
policy makers• Involve stakeholders to avoid strong AI problem• Sensitivity analysis• Game-based Compositional cross-validation
Acknowledgements:• Frank Dignum, Virginia Dignum, Gert-Jan
Hofstede, Tim Verwaart, Saskia Burgers