final presentation
TRANSCRIPT
DOAS FINAL PRESENTATION
Interactive Response System for Crisis Management
4-2-2005
Gijs Dubbelman [[email protected]]Ivo Everts [[email protected]]
Tom van der Weide [[email protected]]Siwei Wang [[email protected]]
Introduction
• Police, firesquad and medics need up to date information during a crisis (via a control room).
• Sensors alone may not fulfill this need, and/or they may be malfunctioning.
• People near the crisis area can provide valuable information.
Our task
• Design and implement a system that can determine whether or not more information is needed about a crisis situation. If so, contact to humans in the crisis area in order to retrieve the desired information.
Distributed Perception Networks
• We use a DPN.
• Recap: A DPN is a multi-agent based approach to fusion of heterogeneous and distributed data. Bayesian inferences can be made. Data fusion node: Fusion Agent, sensor interpreting node: Sensor Agent.
• Actually, we use JavaBayes, which is also being used by the UvA DPN sofware.
Basic system flow (1)
• A sensor registers a sudden change in value on a certain location.
• All Sensor- and Fusion- Agents that are affected by this sensor are investigated, by the Concept Manager Agent.
• Those Fusion Agents with high uncertainty in their hypotheses (see later on), get activated.
• Now the DPN is created.
Basic system flow (2)
• A Human Agent (…) is added to each Fusion Agent.
• This Human Agent contacts the Callcenter and sends a query about the hypothesis.
• The Callcenter sends the query as an SMS to the appropriate human(s), and receives an answer.
• The network gets updated and the control room informed about the evidence.
System designDistributed Perception Network
Concept ManagerAgent
Yellow pagesagent
Callcenter
SMS gateway
EnvironmentSMS SMS SMSSMSSMS
S S S SS S S S S S S S SSMSSMS
Fusion agent
FA HA
SMS Gateway
BD
Concept Manager Agent (1)
• Responsible for DPN initialization.• Bottom up approach.• A sensor is monitoring its environment and sends its
data to its Sensor Agent.• At a certain moment the sensor data may provide
enough evidence for the sensor agent to set its value to true.
DPN
Environment
S S S S S S
SA SA SA SA
FA FA FA
C C
CMA
FA FA
CDPN
Environment
S S S S S S
SA SA SA SA
FA FA FA
C C
CMA
FA FA
C
Concept Manager Agent (2)
• It then sends a message to the Concept Manager.
• The Concept Manager can select all Fusion Agents that have the particular Sensor Agent as one of its possible children.
• An activation message is sent to these agents
DPN
Environment
S S S S S S
SA SA SA SA
FA FA FA
C C
CMA
FA FA
CDPN
Environment
S S S S S S
SA SA SA SA
FA FA FA
C C
CMA
FA FA
C
Concept Manager Agent (3)
• When the agents receive the activation messages they will spawn their own world models, using a top down approach.
DPN
Environment
S S S S S S
SA SA SA SA
FA FA FA
C C
CMA
FA FA
CDPN
Environment
S S S S S S
SA SA SA SA
FA FA FA
C C
CMA
FA FA
C
Concept Manager Agent (4)
• When there is enough evidence to determine with certainty that the agent concept is true then again a message is sent to the CMA.
• The process repeats untill top concepts are reached.
DPN
Environment
S S S S S S
SA SA SA SA
FA FA FA
C C
CMA
FA FA
CDPN
Environment
S S S S S S
SA SA SA SA
FA FA FA
C C
CMA
FA FA
CDPN
Environment
S S S S S S
SA SA SA SA
FA FA FA
C C
CMA
FA FA
C
Intermezzo: Certainty in the system
• When is a concept, represented by a Fusion Agent, uncertain?
• Rephrase: when are we maximally uncertain about a hypothesis?
• … when p(hypothesis) = 0.5
• Concept: ‘fire’
• Hypothesis: ‘there is a fire’
(Un)certainty curve: Uncert(p) = N(0.5, 0.25)(p)
Cert (p) = 1.0 –( Uncert(p) / Uncert(0.5) )Uncertainty above threshold means that
the hypothesis is uncertain
System design cont. Human Agents (1)
• The goal of a Human Agent is to remove uncertainty about the hypothesis of the Fusion Agent to which it belongs.
• Note: it already knows that Uncert(hypothesis) < threshold.
• It can percieve and update the Fusion Agent.
• It can communicate with the Callcenter.
Human Agents (2)
• A Human Agent can undertake one of two possible actions: he can contact the Callcenter, or he can choose not to do so.
DPN
Environment
S S S S S S
SA SA SA SA
FA FA
FA
Callcenter
HA
HA
HA
Human Agents (3)
• Assign a priority measure to a hypothesis: priority(hypothesis)=uncertainty(hypothesis)
• Do not contact if a lot of agents want to contact the Callcenter.
• utility of an agent i to contact the CC: utility(i) = priority(i) + ( 1 / length(queue) )
• Idea: always contact if the hypothesis is maximally uncertain and/or if the HA is the only one in the queue.
Human Agents (4)
HA :: CC Communication
• If a Human Agent’s utility is high enough, it connects to the Callcenter through a socket.
• The Callcenter is continuously listening to a port for requests.
• = Client server application.
• Supports distributed character of the system.
Callcenter (1)
• A priority queue is maintained to store the incoming queries.
• The Callcenter decides which people to contact.
• These people have to be located. We simulated this with a database.
Callcenter (2)
Callcenter
SMS gatewaySMS SMS SMSSMSSMSSMSSMS
Fusion agent
FA HA
SMS Gateway
BD
Callcenter (3)
• In the database also reside human properties: age, proffesion etc.
• Depending on the query, people are selected that are most likely to give a credible response: we prefer a doctor over an AI student for medical questions, and prefer AI students over 10 year olds in general (?).
• If the question is general, a simple broadcast is done to all people in the crisis area
Callcenter (4)
• Responses are also weighted: the response of a gas expert will be weighted heavily for a gas-query.
• If #‘yes’ > #‘no’, send ‘yes’ back to the Human Agent.