1. AFECS Multi-Agent Framework for Experiment Control Systems
V. Gyurjyan, JLAB DAQ Group
4/6/2019
JLAB Computing Seminar

2. JLAB Computing Seminar
4/6/2019
JLAB Computing Seminar

3. JLAB Computing Seminar
Outline
Service Oriented Architecture (SOA)
Agent technologies
Ontology languages
Resource Description Framework (RDF)
AFECS
4/6/2019
JLAB Computing Seminar

4. JLAB Computing Seminar
SOA
Style of building reliable systems that deliver functionality as services
Loose coupling between interacting services
4/6/2019
JLAB Computing Seminar

5. JLAB Computing Seminar
Service
Atomic unit of an SOA.
Encapsulates logic, data, or process.
Autonomous
Location Transparency
It is defined by the messages it can accept and the responses it can give. Messages can be one-way, synchronous or asynchronous.
They can be integrated to provide higher-level services
It can be completely self-contained, or it may depend on the availability of other services or other resources.
4/6/2019
JLAB Computing Seminar

6. JLAB Computing Seminar
Service interaction
Service
Registry
Discover
Advertise
Service
Consumer
Service
Provider
Interact
4/6/2019
JLAB Computing Seminar

7. JLAB Computing Seminar
Loose Coupling
Flexibility
Scalability
Replaceability
Fault tolerance
Service
Registry
Discover
Advertise
Service
Provider
Consumer 1
Data storage 2
Service
Provider 3
Service
Provider
Services are stateless
4/6/2019
JLAB Computing Seminar

8. SOA is for distributed computing
SOA design can be used for a single system - individual processes as services - communication through well defined interfaces using local channels or high-speed interconnect.
4/6/2019
JLAB Computing Seminar

9. JLAB Computing Seminar
SO vs. OO
OO advocates tight coupling, i.e. users have static knowledge of the object.
SOA promotes the design of the system in which it is not necessary that requests and responses are handled by the same set of communicating entities.
Service has an internal thread of control.
sayHelloWorld
Say hello world.
“Hello World” OO SO
4/6/2019
JLAB Computing Seminar

10. JLAB Computing Seminar
What is an agent?
SOA service and much more...
„one that acts or has the power or authority to act... or represent another..“
American Heritage Dictionary.
4/6/2019
JLAB Computing Seminar

11. JLAB Computing Seminar
An agent is
4/6/2019
JLAB Computing Seminar

12. Internal and External Environment of an Agent
Balance
External Environment:
user, other agents,
applications, information sources,
their relationships,
platforms, servers, networks, etc.
Internal Environment:
architecture, goals, abilities, sensors,
effectors, profile, knowledge, etc.
4/6/2019
JLAB Computing Seminar

13. JLAB Computing Seminar
What does it mean?
Intelligent Agent is an entity that is able to keep continuous balance between its internal and external environments (it has at least one goal) in such a way that in the case of imbalance agent can:
change external environment to be in balance with the internal one (creative)
change internal environment to be in balance with the external one (adaptive)
find out and move to another place within the external environment where balance occurs without any changes (mobile)
closely communicate with one or more other agents to be able to create a community, which internal environment will be able to be in balance with the external one (social)
4/6/2019
JLAB Computing Seminar

14. JLAB Computing Seminar
An agent is also
Learning
Trustworthy: an agent will not knowingly provide false information to its user
Benevolent: agents do not have conflicting goals, therefore every agent will always try to do what it is asked for
Rational: agents act in order to achieve their goals and will not act in such a way such as to prevent their goals being achieved
4/6/2019
JLAB Computing Seminar

15. Benefits of the agents systems
Take advantage of distributed computing resources
To coordinate teams of interacting components (each component has physically separate processor and is capable of acting independently)
To increase system robustness
To simplify modeling of complex processes
M. Minsky, The Society of Mind “Human mind is made up of an interacting society of many small agents.”
4/6/2019
JLAB Computing Seminar

16. JLAB Computing Seminar
4/6/2019
JLAB Computing Seminar

17. Standardize interoperability
Language
Semantic
Registration
Discovery
Messaging
4/6/2019
JLAB Computing Seminar

18. JLAB Computing Seminar
FIPA
Foundation for Intelligent Physical Agents
is a non-profit association of companies, founded in Today FIPA counts about 65 member companies.
Its purpose is to promote the success of emerging agent-based applications, services, and equipments
Facilitate the end-to-end networking for heterogeneous and interacting agents and agent-based systems
4/6/2019
JLAB Computing Seminar

19. JLAB Computing Seminar
FIPA specification
Agent Management and platform services
- Agent management services (AMS)
- Directory service (DF)
- Message transport service (ACC)
Agent Communication Model and Language
Set of common Interaction Protocols
4/6/2019
JLAB Computing Seminar

20. JLAB Computing Seminar
FIPA Agent platform
Environment within which agents can be created, registered, located, communicated, migrated, and retired. DF
AMS
ACC
Agent Containers
FIPA Distributed Platform
4/6/2019
JLAB Computing Seminar

21. JLAB Computing Seminar
Directory service
Shared information repository
Agent publish/modify/delete their agent-descriptions including
- Name and address
- Provided services
- Interaction protocols
- Ontology
- Ownership
- Properties
- Etc.
Agents may search for agent description of interest.
4/6/2019
JLAB Computing Seminar

22. JLAB Computing Seminar
ACL
Based on a combination of speech acts, predicate logic and ontology, standard ways of interpreting communication between agents can be achieved that respect the intended meaning of the communication
4/6/2019
JLAB Computing Seminar

23. JLAB Computing Seminar
ACL message structure
4/6/2019
JLAB Computing Seminar

24. Communication protocols
Request
Query
Request when
Contract net
Iterated contract net
English auction
Dutch auction
Brokering
Subscribe
Propose
4/6/2019
JLAB Computing Seminar

25. JLAB Computing Seminar
AFECS ACL
Not understood
Inform-content
Refuse-reason
Act
Agree
Failure-reason
4/6/2019
JLAB Computing Seminar

26. JLAB Computing Seminar
4/6/2019
JLAB Computing Seminar

27. JLAB Computing Seminar
What is an Ontology?
An ontology is a specification of a conceptualization. – Tom Gruber
An ontology is a way of representing a knowledge, meaning that knowledge is formalized in a symbolic form, that is, to find a symbolic expression that can be interpreted. – Klein and Methlie
4/6/2019
JLAB Computing Seminar

28. No really what is an Ontology?
An ontology specifies a description of the
Terminology and concepts
Properties explicitly defining the terms, concepts
Relations among concepts
Rules distinguishing concepts, refining definitions and relations (constraints, restrictions, regular expressions) relevant to a particular domain or area of interest.
4/6/2019
JLAB Computing Seminar

29. Philosophy of Knowledge
Knowledge we have
Knowledge we know we don't know
Knowledge we are not aware we don't know
4/6/2019
JLAB Computing Seminar

30. JLAB Computing Seminar
Formalized Knowledge D1
4/6/2019
JLAB Computing Seminar

31. Domain Ontology A novel and poem are books. A book is a document.
Taxonomic Knowledge
A novel and poem are books.
A book is a document.
Domain Ontology
document
book
novel
poem
4/6/2019
JLAB Computing Seminar

32. Process Ontology A document author is a woman. author document woman
Relational Knowledge
A document author is a woman.
Process Ontology
author
document
woman
4/6/2019
JLAB Computing Seminar

33. JLAB Computing Seminar
Interface Ontology
Process Ontology
author
document
woman
Process Ontology
name
woman
string
Process Ontology
author
document
string
4/6/2019
JLAB Computing Seminar

34. RDF: Resource Description Framework
RDF is best suited to formalize triple structure of the knowledge base.
XML
RDF
Why not XML?
XML gave us a flexibility to define our own tags. Yet, the construct of the XML is a so called duple consisting of tags and values.
RDF record instead is constructed from three components called resource, property and value.
If we look the human language construct we see that the base of the sentence is also triplet. So, RDF is more powerful in defining the knowledge base and is ideal for meta language development.
Value
Resource
Subject
Tag
Predicate
Property
Object
Value
4/6/2019
JLAB Computing Seminar

35. JLAB Computing Seminar
4/6/2019
JLAB Computing Seminar

36. JLAB Computing Seminar
The only thing worse than an experiment you can’t control is an experiment you can.
4/6/2019
JLAB Computing Seminar

37. JLAB Computing Seminar
Project Goal
Develop a framework for building control systems
open architecture
hierarchical
network distributed
Provide a set of tools to deploy:
test and control systems
alarm systems
control visualization systems, etc.
4/6/2019
JLAB Computing Seminar

38. Experiment Control System Requirements
Control components of the experiment.
Full description of the experiment components.
Framework reflects
Fault tolerance and recovery.
Create and deploy finite state machines.
Control data-taking activities by coordinating the operations of the DAQ sub-systems, online software components, slow control system components and third-party software components, as well as industrial control systems
Provide a facilities to describe external control processes within the system, which will in turn encapsulate foreign components (industrial control systems, etc.) and allow already existing components to communicate with those foreign components as familiar ones
Provide facilities, allowing detailed specification of the experiment components/objects, such as their states, actions and associated conditions, etc.
Means to group individual control systems into the control domains, operating in the general experiment control environment. Every control domain, representing experiment subsystem, can operate independently, yet will summarize information for the above level domains or expand actions to the lower levels.
Provide mechanisms to ensure the fault tolerance and problem recovery of the system.
Provide facilities to create and deploy finite state machines.
Provide web interface for alarm reporting, alarm configuration, and control data visualization.
4/6/2019
JLAB Computing Seminar

39. Design Philosophy of AFECS
Pure Java.
Collaborating autonomous agents.
Each agent represents a hardware or software component.
Messages between agents invoke actions.
Agents behave as finite state machines.
Agents communicate with physical components through standard protocols.
4/6/2019
JLAB Computing Seminar

40. AFECS Agent Categories
Normative agents administrative agent
Create and manage all other agents
Environment administration
Supervisor agents Control groups of Component agents
Component agents Represent hardware or software components of the real world control system. S A
4/6/2019
JLAB Computing Seminar

41. JLAB Computing Seminar
Component Agent
Agents mimic the state of the real (physical) component, they can invoke actions which change the physical component state.
AFECS (control) world.
IPC
IPC
IPC
Real world.
HV power supply
Target system
Gas system
4/6/2019
JLAB Computing Seminar

42. Inter-Process Communication Channel IPC
Agents communicate with their associated physical components using range of communication protocols including:
Tcl-DP (legacy protocol)
cMsg (publish-subscribe messaging protocol, JLAB)
EPICS channel access protocol
SNMP (simple network management protocol)
JDBC
OS shell interface
4/6/2019
JLAB Computing Seminar

43. JLAB Computing Seminar
Supervisor Agent
Supervisor agents have discrete states and they respond to messages from other agents (supervisor or component). S
DC HV control
supervisor agent A A A
Agents representing
Individual power
supplies
HV crate 1
HV crate 2
HV crate 3
DC HV crates
4/6/2019
JLAB Computing Seminar

44. JLAB Computing Seminar
Agents Groupings
Physical (AFECS container) A java virtual machine (JVM) is an agent container. Container is a process. Agent is a set of threads in the process.
Logical (AFECS domain) Agents are grouped into virtual clusters or domains according to their functions. Agents in each domain may be visible to other domains. N S A S S S A A A A
4/6/2019
JLAB Computing Seminar

45. JLAB Computing Seminar
Domain
Reduces complexity of large control systems.
Encourages modularity and encapsulation. (only a few agents and their behaviors are visible outside of the domain limits)
Promotes control system hierarchical structure.
Can be redesigned and tested independently, without affecting the rest of the control system. (as long as the behavior of the visible agents remains the same)
Control data-taking activities by coordinating the operations of the DAQ sub-systems, online software components, slow control system components and third-party software components, as well as industrial control systems
Provide a facilities to describe external control processes within the system, which will in turn encapsulate foreign components (industrial control systems, etc.) and allow already existing components to communicate with those foreign components as familiar ones
Provide facilities, allowing detailed specification of the experiment components/objects, such as their states, actions and associated conditions, etc.
Means to group individual control systems into the control domains, operating in the general experiment control environment. Every control domain, representing experiment subsystem, can operate independently, yet will summarize information for the above level domains or expand actions to the lower levels.
Provide mechanisms to ensure the fault tolerance and problem recovery of the system.
Provide facilities to create and deploy finite state machines.
Provide web interface for alarm reporting, alarm configuration, and control data visualization.
4/6/2019
JLAB Computing Seminar

46. AFECS Platform Distribution
Agents are physically distributed
at the configuration time
at runtime S S S N A A A A
4/6/2019
JLAB Computing Seminar

47. JLAB Computing Seminar
Hierarchy of domains
Physical Components
Normative Agents NR NA NC A
CODA ROC S
CODA EMU A
Front-End
Supervisor agent A
EPICS IOC 1 S S A
EPICS CAG
Grand supervisor S A
Trigger soft
AFECS Platform S
ACC
Trigger hard A
Online ANA A
IPC
IPC
IPC
IPC
WEB
GUI/user
4/6/2019
JLAB Computing Seminar

48. Physical Component Integration
Administrative domain and container
AFECS platform NC NR
Description of the control NA
Component
4/6/2019
JLAB Computing Seminar

49. Control System Description
Designing a control system requires precise descriptions of physical components and their software abstractions.
AFECS uses the COOL language for these descriptions.
4/6/2019
JLAB Computing Seminar

50. Control Oriented Ontology Language COOL
Developed using RDF.
Defines an agent model of the experiment.
Describes functionalities of physical components (state machines)
Describes the semantics of control.
ACL message ontology
Can be generated using GUI.
4/6/2019
JLAB Computing Seminar

51. JLAB Computing Seminar
4/6/2019
JLAB Computing Seminar

52. AFECS COOL Specific Tools
GUI using COOL to build configuration files describing control system.
db2cool: converts current CODA config database to COOL config files.
sml2cool: translates state manager language config files to COOL config files.
4/6/2019
JLAB Computing Seminar

53. JLAB Computing Seminar
COOL GUI
4/6/2019
JLAB Computing Seminar

54. JLAB Computing Seminar
COOL Example
<rdf:RDF'
xmlns:cool='
>
<cool:Control rdf:ID="R1ExternalProcess">
<cool:hasComponent rdf:resource="#ROC1"/>
<cool:hasComponent rdf:resource="#RC"/>
<cool:hasOption rdf:resource="#OPTION"/>
</cool:Control>
<cool:Option rdf:ID="OPTION">
<cool:hasCoda2Component>false</cool:hasCoda2Component>
<cool:hasDataFile>/tmp/test.dat</cool:hasDataFile>
</cool:Option>
<cool:Component rdf:ID="ROC1">
<cool:hasIpc>cMsg</cool:hasIpc>
<cool:hasType>ROC</cool:hasType>
<cool:hasName>Roc1</cool:hasName>
<cool:hasPriority>44</cool:hasPriority>
<cool:hasExternalProcess rdf:resource="#PROC1"/>
</cool:Component>
<cool:Component rdf:ID="RC">
<cool:hasType>RCS</cool:hasType>
<cool:hasName>R1ExternalProcess</cool:hasName>
<cool:hasPriority>77</cool:hasPriority>
<cool:Process rdf:ID="PROC1">
<cool:hasCommandIpc>shell</cool:hasCommandIpc>
<cool:isPartOfState>predownload</cool:isPartOfState>
<cool:hasCommandName>emacs</cool:hasCommandName>
</cool:Process>
</rdf:RDF>
4/6/2019
JLAB Computing Seminar

55. AFECS Implementations
New run control for the JLAB data acquisition system. (V. Gyurjyan, et al. “Jefferson Lab Data Acquisition Run Control System”, Proceeding of the CHEP conference. CERN , Volume 1, page 151.)
CLAS experiment web based monitoring system.
4/6/2019
JLAB Computing Seminar

56. JLAB Computing Seminar
Run-Control GUI
4/6/2019
JLAB Computing Seminar

57. JLAB Computing Seminar
Conclusions
Java based framework for designing and implementing hierarchical, distributed control systems with intelligent agents.
Encourages abstraction, encapsulation, and modularity.
Provides a special ontology language (COOL) to describe hierarchical control structure, control logic, and state machines.
Has been successfully used to develop run control system for the JLAB data acquisition system, and CLAS experiment web based monitoring system.
4/6/2019
JLAB Computing Seminar

58. JLAB Computing Seminar
Thank you
4/6/2019
JLAB Computing Seminar