from inter-agent to intra-agent representations
TRANSCRIPT
From Inter-Agent to Intra-AgentRepresentations
8 March 2014 - ICAART presentation
Giovanni Sileno [email protected] BoerTom van Engers
Leibniz Center for Law – University of Amsterdam
mapping social scenarios to agent-role descriptions
Fundamental division: “artificial” social systems norm-driven“natural” social systems norm-guided
In the latter, non-compliance (intentional or not) is systemic.
Example: Humans!
In “natural” social systems, agents do not have a blue-print describing the “implemented” behaviour of other components.
Still, social agents need to understand/interpret, to an adequate extent, how the others behave!
Where knowledge comes from?
In “natural” social systems, agents do not have a blue-print describing the “implemented” behaviour of other components.
Still, social agents need to understand/interpret, to an adequate extent, how the others behave!
How we transmit knowledge about people’s behaviour?
The social function of “stories”
“Many different root metaphors have been put forth to represent the essential nature of human beings: homo faber, homo economous, homo politicus, [...], “rational man”. I now propose homo narrans to be added to the list.”Fischer [1984], Narration as a Human Communication Paradigm
Stories are “constituents of human memory, knowledge, and social communication”Schank and Abelson [1995], Knowledge and memory: the real story
What “stories” are
Not only fictional narrations..
but also:
• personal experiences
• journalistic reports
• medical cases
• legal cases
• … and any other expert domain case!
What “stories” are
Not only fictional narrations..
but also:
• personal experiences
• journalistic reports
• medical cases
• legal cases social behaviours with
their legal interpretation
• …
A relevant issue
Unfortunately, stories, by their own nature, are partial (ill-defined) representations.
What is in a story depends on
• default assumptions
• common-sense knowledge
• expertise
• interest
• focus
• ...
A relevant issue
Unfortunately, stories, by their own nature, are partial (ill-defined) representations.
What is told in a story depends on
• default assumptions
• common-sense knowledge
• expertise
• interest
• focus
• ...
of the narrator !
A relevant issue
Unfortunately, stories, by their own nature, are partial (ill-defined) representations.
What is read in a story depends on
• default assumptions
• common-sense knowledge
• expertise
• interest
• focus
• ...
of the listener !
Our objective
We look for a methodology to acquire the systemic knowledge of the narrator, concerning a given social scenario, in a computational form.
Our objective & requirements
We look for a methodology to acquire the systemic knowledge of the narrator, concerning a given social scenario, in a computational form.
• bypass natural language issueswe are not targeting a story understanding
application, but a scenario acquisition tool
Our objective & requirements
We look for a methodology to acquire the systemic knowledge of the narrator, concerning a given social scenario, in a computational form.
• bypass natural language issueswe are not targeting a story understanding
application, but a scenario acquisition tool
• target non-IT experts (in principle) we will refer mostly to diagrams, or programming
based on high level and “intuitive” languages
affinity with scenario-basedmodelling
A very “simple” story
A seller makes an offer, about a certain good, for a certain amount of money. A buyer accepts. The buyer pays. The seller delivers.
Outline of the methodology
1. start from an inter-agent description
2. enrich it with intentional/institutional factors
3. synthetize it in intra-agent models
Inter-Agent Description: MSC
Message Sequence Charts (MSC) formalize UML sequencediagrams. They are intuitive and commonly used.
Inter-Agent Description: MSC
Issues with:
• “ontological” identities
• Neglecting side-effects/control-loop
• implicit ordering
Inter-Agent Description: Topology
Identity is epistemic (assigned to message boxes)
Inter-Agent Description: Flow
Problems:
• Consecutiveness vs Consequence“the mainspring of the narrative activity is to be traced to thatvery confusion between consecutiveness and consequence, what-comes-after being read in a narrative as what-is-caused-by”, Barthes [1968]
• Story-relative vs Discourse-relative timelinesordering as events occur or how they are told/observed
Inter-Agent Description: Flow
Three levels of constraints on events:
• dependencies (logic or causal)
• relative/absolute time indexing
• discourse ordering
The first is by far the most important to our scope.
An important step is the recognition of sub-systems operating concurrently (e.g. agents, cognitive modules)
Inter-Agent Description: Flow
Outline of the methodology
1. start from an inter-agent description
2. enrich it with intentional/institutional factors
3. synthetize it in intra-agent models
First steps toward an agenticcharacterization
• Intentional characterizations what the agents want?
• Hidden acts is there something else that they perform?
(e.g. evaluations, information retrieval)
• Critical conditions
is there any condition necessary for the performance
and the effectiveness of an action?
Agentic Characterization on MSC
Agentic Characterization on MSC
Issues:
ontological–epistemic
emission–reception
Agentic Characterization: Refinement
• Task decomposition per agent(additional independent flows)
Agentic Characterization: Refinement
• Task decomposition per agent(additional independent flows)
• Communication synchronization(chaining task decompositions with the story)
• Pragmatic interpretation of messages (via Speech act theory), e.g. a promise is a
commitment and generates an obligation
Agentic Characterization: Refinement
• Target: multi-layered representation
Main component
Signal layer Message / Act
Action layer Action / Activity
Intentional layer Intention
Motivational layer Motive
Agentic Characterization: Refinement
• Target: multi-layered representation
Main component
Signal layer Message / Act
Action layer Action / Activity
Intentional layer Intention
Motivational layer Motive
• Motives are reasons for action
• Obligations usually are prototypical motives.
Agentic Characterization: Refinement
• Target: multi-layered representation
Main component Catalyser
Signal layer Message / Act
Action layer Action / Activity Disposition
Intentional layer Intention Affordance
Motivational layer Motive Motivation
• Affordance: perceived contextual power
• Disposition: actual contextual power
also for the institutional domain
Agentic Characterization on Petri Net
• Multi-layered
• Events / Conditions places
• Synchronization on message layer
Outline of the methodology
1. start from an inter-agent description
2. enrich it with intentional/institutional factors
3. synthetize it in intra-agent models
Intra-agent synthesis in AgentSpeak(L)/Jason (example)
+!pay_to(Amount, Agent)
: owning(Money) & Money >= Amount
<- .send(w, achieve, pay_to(Amount, Agent));
+paid_to(Amount, Agent).
• Such scripts give only internal and epistemic perspective.
• Synchronization and ontological factors to be implemented in environment w
Conclusions & further developments
Multiple interpretations of the same story are possible. As far as they produce the correct messages they are valid models.
Each representation (MSC, Topology, Petri Nets or AgentSpeak(L)/Jason scripts) have its own pro/cons. An adequate integrated environment should allow to pass from one to the other.
Necessity of defining operators of “distance” and “subsumption” to compare/integrate stories.
Conclusions & further developments
Scenarios acquired through this methodology can be collected, furnishing a deep model of a social setting model-based diagnosis
Alternatively, they can be executed on a simulation engine, in order to test new policies/regulations environmental models for a design tool
Conclusions & further developments
Multi-Agent Systems research and practice usually target “artificial” social systems.
The closure of the system comes by design or as strict assumption
basis for all analytical tools
guidance != controlas institutions influence agents, agents influence institutions a constructivist approach toward MAS
Conclusions & further developments
Multi-Agent Systems research and practice usually target “artificial” social systems.
The closure of the system comes by design or as strict assumption
basis for all analytical tools
guidance != controlas institutions influence agents, agents influence institutions a constructivist approach toward MAS
Intra-agent synthesis in AgentSpeak(L)/Jason (example)
+!accept(offer(Good, Amount)[source(Seller)])
<- .send(Seller, tell, accept(offer(Good, Amount)));
+obl(pay_to(Amount, Seller)).
+obl(pay_to(Amount, Agent))
<- !pay_to(Amount, Agent);
-obl(pay_to(Amount, Agent)).
+!pay_to(Amount, Agent)
: owning(Money) & Money >= Amount
<- .send(w, achieve, pay_to(Amount, Agent));
+paid_to(Amount, Agent).
• Such scripts give only internal and epistemic perspective.
• Synchronization and ontological factors to be implemented in environment w