SLIDE 1
Open Reflective Agents Sylvain Giroux LICEF, Tl-Universit 1001 - - PDF document
Open Reflective Agents Sylvain Giroux LICEF, Tl-Universit 1001 - - PDF document
Open Reflective Agents Sylvain Giroux LICEF, Tl-Universit 1001 Sherbrooke est C.P. 5250, succ. C Montral, P.Q. Canada H2X 3M4 sgiroux@teluq.uquebec.ca PLAN 1- Motivation and approach 2- Adaptation 1- The viewpoint of Gregory
SLIDE 2
SLIDE 3
2
Motivations The Problem
- Real systems are open ---> Agents ought to be open.
- When an agent is intended for open multi-agent systems,
its designer can
- > neither exactly predict behaviour of other agents
- > nor rely on some sort of Esperanto.
- Stating the problem differently, interoperability is an issue ensuing from
- penness.
Issues raised by openness and directions towards solutions #1 how agents can interact despite of different behaviours and models of computation ? agents should at least adhere to a common ontology #2 how could agents live and act in evolving and unpredictable worlds? they must have adaptive mechanisms Long-term benefits expected from open agents
- Agents will learn to interact together,
so designers and users interventions are less likely to be needed.
- Open agents increase the robustness
- f both agents and multi-agent systems.
- Agents and multi-agent systems will become easier to design and maintain; for
instance, to build up a system, one just has to hire pre-built open agents.
SLIDE 4
3
Our Approach
1- Definition of a minimal ontology that could set the guidelines for open adaptive agents. 2- ReActalk as a reflective platform for the study of adaptation in open multi-agent
- systems. ReActalk helps
- viewing information systems as open agents
- viewing agents as open information (eco)systems
- viewing interoperabilty as an adaptive process
3- Application of ReActalk architecture and adaptive mechanisms a) to solve interoperability issues arising around Actalk actors b) to build an open agenda system c) EpiTalk, epiphyte information systems applied to advisors d) SMAVOR, volcanoes models and simulation
SLIDE 5
4
Adaptation The viewpoint of Gregory Bateson
- As a starting point, we adopted the viewpoint of Bateson on adaptation
“adaptation : A feature of an organism whereby it seemingly fits better into its environment and way of life. The process of achieving t hat fit. ” [Bateson, Mind and Nature]
- Remarks
1- Bateson underlines the relation that links together a living being and its environment a living being and its evolution cannot really be understood unless the living being is understood as part of a larger system 2- There are merely two ways to adapt
- --> to self-modify
example: speeding up one's heart rate at a high altitude
- --> to modify its environment
example: agriculture
SLIDE 6
5
On Mechanisms for Adaptation
- to assemble and modify dynamically models of computation in respect to
interactions an agent has with its environment
- > conscious adaptation
analogy: learning from social intercourse purpose:
- to solve interoperability issues
requirements:
- to detect situations calling for adaptive modifications
- to trigger adaptive modifications
- to carry out adaptive modifications incorporating new abilities
- > unconscious adaptation
analogy: forgetting purposes:
- not to overweight needlessly the agent internal processes
- to keep the sole abilities relevant in the current ecosystem of the agent
requirements:
- to decide between obsolete abilities and helpful ones
- to trigger adaptive modifications
- to remove obsolete abilities
SLIDE 7
6
From Real World towards Software Agents
adapt at ion env ironment Art if icial Life
- t o reason and act upon oneself
Dist r ib ut e d A r t i f ic ia l Int elligence Open Syst ems (Carl Hewit t )
- concurrency
- asynchronous
- decent ralized cont rol
Sm allt alk - 8 0 but class <=> umbilical cord passive and sequent ial object s Re f le ct ion Re A c t a lk act ors dynamic and t emporary modificat ions of oneself
- n a individual basis
- modelling
- programming environment
- from isolat ion
t owards a syst emic approach agent s
- how global macro-behaviors
do emerge out of local micro-behaviors
- ecosyst em reificat ion
but laws regulat ing t he world are kept occult but an act or does not consider it is embedded int o a syst em but art ificial beings usually lack mechanism enabling t hem t o modify t heir behavior
- b ject s
SLIDE 8
7
The Viewpoint Embodied within ReActalk
- the world is made of actors that are active and concurrent;
- an actor is reflective and thus can reason or act upon its own behavior;
- an actor can always be decomposed into an organisation of actors which acts
as its reflective representation;
- an agent is an actor, member of an organisation;
- an ecosystem is reified as an organisation,
that is through the vantage point of its members and relations amongst members;
- interactions are performed through message passing.
SLIDE 9
8
An Actor Internal Adaptive Behavior
Basis: Actalk
- small kernel
- already available extensions towards actor models of computation
- to take advantage of Smalltalk-80 programming environment
Actalk was developed by Jean-Pierre Briot Objective: to render explicit the computational behavior of an actor
- > in order to modify it dynamically
Observation: an actor behavior is driven by messages (external stimuli) Analogy: computational behavior <----> factory messages <----> pieces actors / agents <----> workers How: by isolating each step of the processing of a message
SLIDE 10
9
A Generic Factory
arrivalOf: aMessage
MetaActor
ArrivalManager Receptionist MailboxOrganizer MailBoxRetreiver ScriptManager
after execution execution of script before execution
ExecutionManager d e n
- ta
t i
- n
meta (as an agent) aMessage
m a i l b
- x
b e h a v i
- r
ReflectiveActor
meta (as an individual)
ReflectiveActorBehavior
a s e l f agents = {ArrivalManager, ... } relations ={ (ArrivalManager, Receptionist, CommunicationProtocol), ... } d e n
- t
a t i
- n
Organization
meta (as an agent)
Organization
SLIDE 11
10
On Mechanisms for Adaptation
- to assemble and modify dynamically models of computation in respect to
interactions an agent has with its environment
- > conscious adaptation
analogy: learning from social intercourse purpose:
- to solve interoperability issues
requirements:
- to detect situations calling for adaptive modifications
- to trigger adaptive modifications
- to carry out adaptive modifications incorporating new abilities
- > unconscious adaptation
analogy: forgetting purposes:
- not to overweight needlessly the agent internal processes
- to keep the sole abilities relevant to the agent
current ecosystem requirements:
- to decide between obsolete abilities and helpful ones
- to trigger adaptive modifications
- to remove obsolete abilities
SLIDE 12
11
Conscious Acquisition of Somatic Abilities
rest ructuration t hrough hy br idiz at ion
SLIDE 13
12
On Mechanisms for Adaptation
- to assemble and modify dynamically models of computation in respect to
interactions an agent has with its environment
- > conscious adaptation
analogy: learning from social intercourse purpose:
- to solve interoperability issues
requirements:
- to detect situations calling for adaptive modifications
- to trigger adaptive modifications
- to carry out adaptive modifications incorporating new abilities
- > unconscious adaptation
analogy: forgetting purposes:
- not to overweight needlessly the agent internal processes
- to keep the sole abilities relevant to the agent
current ecosystem requirements:
- to decide between obsolete abilities and helpful ones
- to trigger adaptive modifications
- to remove obsolete abilities
SLIDE 14
13
Removing Obsolete Somatic Abilities and Metalevels
- behavior
reif icat ion def ault model of comput at ion
- rganizat ion
reif icat io n rest ruct urat ion t hrough unhy bridiz at ion behavior reificat ion hyb rid model of comput at ion (2) r e v e r se surveillance
- n
the met alevel (1) (3)
SLIDE 15
14
Conclusion
- In order to adapt, an agent modifies itself at the mercy of interactions within
its ecosystem. These interactions trigger its adaptive mechanisms.
- Organizational reflection, thanks to adaptive mechanisms, allows to smoothly
assemble on the fly multi-agent systems and to intertwine agent specifications, easing programs cooperation and consequently the programmer tasks.
- Two identical agents inserted into different ecosystems can experience
divergent evolution and may later exhibit distinct behaviors. Reflection frees an agent from its genetic behavior (defined by its class) and gives it the keys to somatic evolution.
SLIDE 16
15
Esquisse d’une ontologie
A1 le monde réel est composé d’objets; A2 les systèmes du monde réel sont fondamentalement ouverts; les systèmes réels sont reliés au monde externe et communiquent avec lui; A3 le monde réel évolue en parallèle; A4 toute entité appartient à un environnement elle doit être comprise en relation avec cet environnement; A5 toute entité a besoin de s’adapter à des situations nouvelles. par conséquent, elle doit pouvoir soit se modifier dynamiquement, soit modifier dynamiquement modifier son environnement
SLIDE 17
16
SAGE : un système d’agendas I
- Adaptation
- bjectif :
mise en évidence des concepts liés à l’adaptation
- chaque agent lié à un agenda possède sa propre culture
culture = modèle de calcul
- chaque agent doit s’adapter à la culture des autres
technique développée --> hybridation
- écosystème des agendas doit pouvoir évoluer dynamiquement
>>> ajouts et retraits dynamiques d’agendas >>> ajouts et annulations dynamiques de rendez-vous
- --> distribution du processus de prise de rendez-vous
SLIDE 18
17
SAGE : un système d’agendas II
- Programmation par écosystèmes
- bjectif :
vérifier la pertinence de la programmation par écosystèmes pour la construction de systèmes complexes
- chaque usager peut émettre des politiques quant à la prise de rendez-vous
politique = loi régissant l’écosystème agenda
- les agents liés à un agenda essaient de se conformer aux politiques émises tout
en tenant compte des circonstances
- -> à travers les choix envisagés dans la prise de rendez-vous
- tout usager peut intervenir sur la prise de rendez-vous
- -> distribution du processus de prise de rendez-vous
SLIDE 19
18
Du besoin d’une ontologie les systèmes adaptatifs sont en interaction avec le réel les systèmes informatiques le monde physique Quelle est la situation actuelle ? si les systèmes à construire sont de la même nature que les systèmes avec qui ils interagissent intégration plus facile uniformité Pourquoi une ontologie ? la réalité est source inspiration contraintes Que faut-il en conclure ?
SLIDE 20
19
Intégration de ReActalk et NéOpus Pourquoi ?
- représentation déclarative
édicter des politiques
- tester la généralités des mécanismes adaptatifs
en particulier, l’hybridation
- sortir des univers à acteurs
pour véfifier la généralité des idées sous-jacentes à ReActalk
- introduire le parallélisme et le dynamisme dans NéOpus
Adaptation consciente dans ReActalk
- réactive : provoquée par les stimuli
- dynamique
- provoque une ascension vers les métaniveaux
Représentation explicite du contrôle dans NéOpus
- décidée au départ lors de l’association de la métabase et de la base
- statique
- descendante : de la métabase vers la base de règles
Solution : reprendre les mécanismes adaptatifs de ReActalk
- réification paresseuse des métaniveaux
- (dés)hybridation pour modifier le comportement
SLIDE 21
20
NéOpus
Intégration d’un mécanisme d’inférence d’ordre 1 (variables) en chaînage avant dans Smalltalk-80 Les règles sont des méthodes Les bases de règles sont des classes ---> des objets Smalltalk-80 Une règle NéOpus peut filtrer n’importe lequel objet Smalltalk-80
- --> en particulier, les bases de règles
>>> expliciter le contrôle à l’aide de base de méta-règles
- n peut associer à volonté une base de règles et une base de méta-règles