© 2012 IBM Corporation © 2012 IBM Corporation DEBS 2012 presentation: A basic proactive model Yagil Engel, Opher Etzion , Zohar Feldman IBM Haifa Research Lab
May 09, 2015
© 2012 IBM Corporation© 2012 IBM Corporation
DEBS 2012 presentation: A basic proactive model
Yagil Engel, Opher Etzion, Zohar Feldman
IBM Haifa Research Lab
© 2012 IBM Corporation2
What are we trying to achieve?
The basic proactive model is applicable for certain types of applications, it is a first phase in building a library of proactive models
“Rapid business, economic, social, and political changes are leading organizations to shift their thinking from reactive (sense and response) to proactive (seek, model, and adapt) in order to detect opportunity and threat events that could affect their business”.
Gartner #208030, December 2010
The goal is to apply the right action at the right time to gain optimal value for a quantitative metric, given an anticipated unplanned event .
© 2012 IBM Corporation3
Some features of the problems we are approaching
There is a quantitatively significant value of mitigating/preventing anticipated event. the goal is to optimize this value
The way to anticipate the event is by itself event-driven (causality relations among events, or situation driven activation of prediction model), the events may have some uncertainty associated with them
The space of possibilities is too large and it is not feasible to compute all states offline
The timing of detection and of action can change the results – decision and action have real-time constraints
The anticipated event is uncertain, and its occurrence time is also uncertain – the prediction contains occurrence time expectancy over a relevant time interval
© 2012 IBM Corporation4
Let’s start with a simple story
An oil drilling session started in February 1st 6:00 and is scheduled to last until February 11th 18:00
There are variety of sensors checking various factors that might cause equipment break – for the story we’ll concentrate on a single one: surface temperature
The monitored pattern is “surface temperature is consistently at least 4% more than upper limit for a period of 10 minutes”
Temporal context: overlapping sliding window of 10 minutes from each measurementSegmentation context: surface
Pattern: For each measurementtemperature > 1.04 *
surface.upper_limit
When detecting this pattern we are interested in knowing: when a crash is expected (and how likely is it)? what is the best action from cost/benefit perspective given: time of detection, expected time of crash, duration to end of the drill, available options
© 2012 IBM Corporation5
When is the crash expected? Temporal context: overlapping sliding window of 10 minutes from each measurementSegmentation context: surface
Pattern: For each measurementtemperature > 1.04 * surface.upper_limit
We would like this pattern to generate a derived event called “equipment crash” whose occurrence time is in the future
The timing of the crash event is uncertain, it is expressed as EXPECTANCY DISTRIBUTION OVER THE TIME INTERVAL BETWEEN NOW AND DRILL END
Online information:Detection time, sizeof interval, trend of Temperature measurementsince start of drill
Prediction model is createdoffline using regularprediction modeling.
Anevent
pattern
NOW Drill end
1
0
Feb 8, 10:00
80
Feb 11, 18:00
© 2012 IBM Corporation6
What are the possible actions?
Lubrication+ low cost; + does not harm productivity;- relative low probability to prevent crash
Operating in low pressure
+ low setup cost; - harms productivity? medium probability to prevent crash
Full maintenance
- high cost - productivity is substantially harmed+ high probability to prevent crash
Questions
1. What is the action that will maximize the utility?
2. When is the best time to activate this activity?
A function of the costs and durations of actions, impact
on the target event
© 2012 IBM Corporation7
Some concrete (simulated) results
Feb 8, 10:00 Feb 11, 18:00
1
The event pattern has been detected in Feb 8, 10:00
Time = 0
2
Normalizing all to cost units – calculation of expected cost distribution for every action was done
)Time =0(
4
The action which minimizes the cost is maintenance at time = 30
3
Feb 9, 16:00
Action:
Schedule maintenance for Feb 9, 16:00
Cost
© 2012 IBM Corporation8
Note that the decision is sensitive to timing of detection
If the detection is done close to beginning of drilling session–
Feb 1, 08:00, then it is better to do lubrication now
If the detection is done closer to the end of the drilling session - to beginning of drilling session – Feb 9, 16:00 then it is better to go to low pressure mode after 30 hours
)Feb 10, 20:00(
Cost
Cost
© 2012 IBM Corporation9
Some experimental results with various scenarios with variance in temperature trends
Y axis = temperature percentage above normal
Myopic = execute the decision now
In scenarios 1 and 3 there are significant improvements when timing of action is also a decision
© 2012 IBM Corporation10
Let’s view some of the characteristics of this example
PropertyOur approach Alternatives
What triggers actionable decision?
Predicted eventRequest, periodic calculation
How is the target event predicted?
Event pattern determines the event, timing and attributes of events by predicting model using event patterns results as input
Pre-calculated, by applying predictive model on request
When is the prediction done?
When the pattern is matchedIn off-line, on request, as part of periodic calculation
When is the predicted event expected to occur?
Over an interval with expectancy distribution
In fixed-time point, somewhere in an interval
How is the decision done?
By a decision process that takes the time distribution of predicted event , costs and duration of actions, expected impacts of actions
By using pre-determined rules, by using pre-determined scoring model, by simulation
When is the action scheduled to be activated?
In the time on which the expected utility is
optimized – part of the decision process.
Immediately when model is applied, by manual decision.
© 2012 IBM Corporation11
Some alternative and complementary approaches
Alternative approach
Pros Cons
Off-line optimization
Generic, good results – can complementour solution as the “typical case”
Low level abstractions, not suitable for real-time
Using rule-based decision
Intuitive, suitable when trade-off is not involved or trivial – can complement our solution to fine-tune the action
Decisions are designed by user, not optimized, not applicable for large number of occurrences.
Sequential decision models (e.g. MDP)
Optimized, considering all possible statesComplementary – adapted version
Complicated, applicable to small amount of states
Reinforcement learning
General, continuously adapted, does not require much modeling
Results may not be optimized, requires significant amount of historical data
© 2012 IBM Corporation12
The proactive use pattern
© 2012 IBM Corporation13
What are the additions to the event processing model?
Forecasted derived events with uncertainty
Introducing proactive agent to the event processing
network
© 2012 IBM Corporation14
Forecasted derived events
In event processing systems derived events are VIRTUAL EVENTS that are assumed to happen when created
In our model forecasted derived events are OBSERVALE EVENTS that are assumed to happen in the future.
The actual occurrence of the event as well as the occurrence time are uncertain and require the extension of the event processing model with uncertainty representation and handling
© 2012 IBM Corporation15
Action*{Ce, Ca(t), d, e’(T)}Action*{Ce, Ca(t), d, e’(T)}
ContextContext
ProducerProducer
ActuatorActuator
Event Processing Agent
Event Processing Agent
ConsumerConsumer
Event Type{name, attribute*}
ProactiveEPA
ProactiveEPA
Forecasted Event Type{name, attribute*, e(T)}
Action {t, parameter*}
Time to take the action
Time distribution of the occurrence of the event until time T - (life expectancy)
Ce – cost of the event if this action is takenCa(t) – cost of the action if taken based on the time it is takend – duration of the actione’(T) – time distribution of the event if action is taken
15
The enhanced event processing model with proactive agents
© 2012 IBM Corporation16
Monitoring of location, time, and magnitude of earthquake, and reported damages
Based on seismic sensors and citizen reports
Monitoring of location, time, and magnitude of earthquake, and reported damages
Based on seismic sensors and citizen reports
Forecasting that within the next 3 hours there will be a a potential damage in a certain location based on an event causality model
Forecasting that within the next 3 hours there will be a a potential damage in a certain location based on an event causality model
Taking proactive actions in notifying and performing actions such as: close roads, stop trains, turn off gas and water supply, evacuate people…
Taking proactive actions in notifying and performing actions such as: close roads, stop trains, turn off gas and water supply, evacuate people…
detect forecast decide act
Real-time decisions about steps and protocols to be followed
Real-time decisions about steps and protocols to be followed
Scenario 1: Disaster management scenario
Scenario properties:Big variance in disaster related developing scenario. Type of decisions vary among casesAspects: life saving, economic, environmental
© 2012 IBM Corporation17
Scenario 2 - Road management scenario
Detect
Monitoring streams of events from sensor in highway and leading ways, from mobile devices, and from accidents reports
Monitoring streams of events from sensor in highway and leading ways, from mobile devices, and from accidents reports
Forecast Act(proactive)
Forecasting that at some point in 10-15 minutes a traffic congestion of certain size will occur in probability of 0.6
Forecasting that at some point in 10-15 minutes a traffic congestion of certain size will occur in probability of 0.6
Taking proactive actions in setting up entry and exit traffic lights durations and speed limit in highway segments
Taking proactive actions in setting up entry and exit traffic lights durations and speed limit in highway segments
Decide(RT)
Scenario properties:Traffic can have chaotic behavior. Amount of possible solutions is very large and requires optimization based on the current observations under strict time constraintsAspects: economic, quality of life, environmental
© 2012 IBM Corporation18
Example 1: Intelligent business operation in surgery rooms (reported by Jim Sinur, Gartner)http://blogs.gartner.com/jim_sinur/2012/01/10/success-snippet-intelligent-business-operations/#comments
The scenario: PREPROCESS - Simulation-based optimization of scheduling and resource allocation off-line for all surgeries planned for the next day
DETECTReal-time tracking of everything: physicians, nurses, equipment; monitor of procedure duration and status - using sensors, cameras - exploiting the "Internet of Things“
FORECASTDetermination of things already going wrong (not according to plan) and anticipation when the surgery will end/resources will be used
ACTRe-applying the simulation based optimization (this time online!) and get updated resource allocation plan.
Scenario 3 - Surgery room scenario (decision by event-based optimization)
Scenario properties:Large variance in behavior of surgeries. There is a need to anticipate and schedule resources (rooms, physicians, equipment)Aspects: life threat, quality of life, economic
© 2012 IBM Corporation19
Scenario 4: merchandise delivery scenario (decision by event-based optimization)
Example 2: Freshdirect (reported by Timo Elliot, SAP)http://smartdatacollective.com/timoelliott/45868/2012-year-analytics-means-business?ref=node_other_posts_by
The scenario:PREPROCESS - Plan distribution of merchandise by trucks
DETECTReal-time tracking of trucks
FORECASTDetermination that in the next hour deliveries planned will be below target
ACTThe company applies its reserve trucks to replace trucks that are behind their schedule and re-plan
Scenario properties:Large variance in travel time, especially in urban areas. Substantially reduce late delivery.Aspects: economic, reputation
© 2012 IBM Corporation20
Summary: what did we achieve? What are the further challenges?
1. The basic proactive model is a feasibility demonstration point for the proactive event-driven paradigm
2. The model built is applicable for a set of applications with specific characteristics
There are a lot of challenges:
Real-time optimization models for other cases
Forecasting models
Consumability by users
Scalability issues