Toward Models of Surprise Eric Horvitz Microsoft Research

Toward Models of Surprise Eric Horvitz Microsoft Research & University of Washington

Surprise Unexpected or umodeled future event or outcome of significant consequence. u Unexpected event: “We thought about it, but considered it would occur with probability < x” u Unmodeled event: “We never even thought about that…” u Significant consequence: Significant positive or negative swing in utility of outcome for observer or group u Events can be personal, local, national, world

Surprise Unexpected or umodeled future event or outcome of significant consequence. u Surprise defined in terms of an observer’s expectations u Deliberation and modeling can influence degree of or nature of surprise Ø Moving from surprise to plausible expectation Ø Expectation of classes or properties of surprise should given that a surprise occurs.

Forecasting Surprise Qualitative and quantative approaches u Qualitative brainstorming about likelihood and nature of surprises over future time periods from knowledge, trends, and historical data u Opportunity: Statistical models of surprise Ø u Learn from databases of past surprise to predict future surprises Opportunity: Statistical models of predictive competency Ø Context- and content-centric failure to predict with accuracy

Surprise. Casting: Abstract and Project u Example: Ø Identify top n most surprising things in prior 1 year (world events or personal life) Ø Repeat for 5 years of surprises Ø Characterize nature and rate of significant surprises Ø Consider top n over longer-term periods in same manner Last 2004 12 months Top n Properties Top n 2003 Properties Top n 2002 Properties

Surprise. Casting: Abstract and Project Ø Turn to the future with statistics about types, examples. Ø Consider triaging of attention and recommending actions. Surprise Forecasting: 2005 -2020 Surprise Types Examples Attention & Information Top n surprise types, by class or properties Assume type occurs; give top m examples per type Recommended triage of focus of attention: additional study, info. gathering, monitoring per type Surprises in technical sophistication of weaponry available inexpensively and universally Largescale simultaneous & continuous disabling of communication and sigint satellites available to multiple states. Actions Recommended actions to react, contain, disrupt negative suprises and/or exploit positive surprises

Surprise. Casting: Abstract and Project Ø Deliberation about potential surprises: “Backcasting” Ø Chaining from surprising outcome to explore feasibility Surprise Forecasting: 2005 -2020 Surprise Types Examples Attention & Information Top n surprise types, by class or properties Assume type occurs; give top m examples per type Recommended triage of focus of attention: additional study, info. gathering, monitoring per type Exotic transnational actors with access to increasing technical prowess. Terrorist organization establishes secret alliance with seemingly friendly former adversary state, intolerant of US “empire”; gains access to nuclear arms with deniability former adversary. Actions Recommended actions to react, contain, disrupt negative suprises and/or exploit positive surprises

Expecting Surprises u Technical developments u International political terrain

Technical Surprises from Past u Big surprises over 50 years Ø Flight Ø From likely unachievable dream, to major pillar of force projection and deterrence in just a “few summers”… Ø Quickly melted into background of obviousness Ø In a few years: From awe of a “canvas flying” contraption to concerns about low-salt meal while streaking across ocean in widebody jet. Ø Nuclear fission and fusion Ø Theory of Relativity and implications out of the blue Ø Shifts in balance of power Ø Risk to nation, human civilization

Classes of Future Technical Surprises? u Computation Ø u Intelligent systems change world (consumer, defense, communications, etc. ) in a fundamental way? Energy Ø Ø Changes nature of powerplants in our and adversaries weapon systems Ø Shift in balance of power hinge on the new source Ø u New surprise source of energy changes economics of world Asymmetric force have new tool on their side for unleashing chaos and disruption Biology Ø Assembly of Polio virus via internet and mail order in 2002 heralds era of bio-hacking Ø Major disruptions via pandemics

Futurist Community: “Wildcards” “A wild card is a future development or event with a relatively low probability of occurrence but a likely high impact on the conduct of business” BIPE Conseil, Copenhagen Institute for Futures Studies, Institute for the Future: Wild Cards: A Multinational Perspective, Institute for the Future 1992

Futurist Community: “Wildcards” u Futurist discussion of wildcards, “futurequakes” Ø Are natural, human-caused, or combinations Ø Are negative and positive Ø Multiple wildcards can interact Ø Breaks, discontinuities vs. trends Ø Peterson (2000), Out of the Blue: How to Anticipate Big Future Surprises Ø Cornish (2003), The Wild Cards in Our Future Ø Steinmüller (2003), The Future as Wild Card: A Short Introduction to a New Concept

On Expected Surprises… u Earth & Sky u Biomedical Developments u Geopolitical & Sociological changes u Technology & Infrastructure Upheaval u Surprise Attack u Spiritual & Paranormal J. Petersen, Out of the Blue: How to Anticipate Big Future Surprises, 2000.

On Expected Surprises… We are now living in another period of significant transition--a foreshortened span of time, during which our surroundings and experiences will change more than during any era in history. Humanity has never lived through the convergence--and, in some cases, the collision--of global forces of such magnitude and diversity. … One foreseeable outcome might be global instability; another, a planetary renaissance. J. Petersen, Out of the Blue: How to Anticipate Big Future Surprises, 2000.

On Expected Surprises… … In any case, during the next two decades, almost every aspect of life will be fundamentally reshaped. . J. Petersen, Out of the Blue: How to Anticipate Big Future Surprises, 2000.

Technological Prowess and Wildcards Rapidly growing technological prowess of humans has, for the first time in history, produced new classes of wild cards that could potentially destroy the whole human race. The new kinds of wild cards are simply too destructive to allow to happen and therefore require active attempts to anticipate them and be proactive in dealing with their early indicators. "Peterson, Out of the Blue: How to Anticipate Big Future Surprises, 2000

Preparing for Wildcards Wild cards can be anticipated and prepared for. Thinking about a wild card before it happens is important and valuable. The more that is known about a potential event, the less threatening it becomes and the more obvious the solutions seem. "Peterson, Out of the Blue: How to Anticipate Big Future Surprises, 2000

Backcasting Cornish: Multipronged Aerial Terrorist Attack was identified, e. g. , 1987 article (B. Jenkins); 1994 article (M. Cetron): "Targets such as the World Trade Center not only provide the requisite casualties but, because of their symbolic nature, provide more bang for the buck. ” "In order to maximize their odds for success, terrorist groups will likely consider mounting multiple, simultaneous operations with the aim of overtaxing a government's ability to respond, as well as demonstrating their professionalism and reach. " “Despite all this, terrorism will remain a back-burner issue for Western leaders as long as the violence strikes in distant lands and has little impact on their fortunes or those of their constituents. ” E. Cornish, The Wild Cards in Our Future. The Futurist. Washington: Jul/Aug 2003. Vol. 37, Iss. 4.

Backcasting How might we proceed to evaluate this warning? u Poll experts on likelihood and targets of aerial suicide attacks. Assume aerial attacks plausible. u Backcasting: Reason about the challenges and solutions. Terrorists are unlikely to have their own aircraft. Could they buy one? Perhaps, but the cost would be enormous and authorities might ask too many questions. Could they hijack a military bomber and load it with bombs? They would have to be awfully clever and lucky to get onto an air base, seize a plane, load it with bombs, and fly it to a target. Success would be unlikely, and terrorists would know it. Could they hijack a military bomber and load it with bombs? Could they steal a plane? § An individual might have difficulty intimidating passengers and crew, but a group could probably do it even with modest weaponry. Getting into the pilot's cabin seems possible. So taking command of the plane seems feasible. But would the pilot follow instructions? Some instructions, maybe, but perhaps not all-especially if asked to do something suicidal. Could a terrorist be trained as a pilot to take over? Yes; a number of men in terrorist organizations have been trained as pilots. But how can an aircraft cause destruction if it has no bombs to drop? It can simply crash into its target. Such a crash should wreck a building; a skyscraper would be especially vulnerable. Of course, everybody on the plane would be killed, but if the terrorists are ready to die themselves, the deaths of the passengers serve as an added benefit. Scores of terrorists have gone on similar suicide missions by strapping bombs to themselves. We can now envision a group of terrorists hijacking an airliner and crashing it into an enemy target. But what target might it be? If we had read Cetron's article, we would already have one target to think about-the World Trade Center. In fact, Islamic terrorists did attack the Center in 1993, but their car bomb failed to destroy the building. §However, Cetron suggested that multiple targets might be selected for an attack. So we might stress the possibility of a larger, wider attack involving more terrorists and multiple planes and targets.

Research on Statistical Models of Surprise u Most reasoning about surprise qualitative u Challenge: Data sparcity u Pushing on theoretical foundations of surprise u Example: Research traffic prediction service, Seattle

Streaming Intelligence Continual sensing, learning, and reasoning always available in stream of daily life Learning & reasoning Sensing Web service Portable device Rich sensing

Predictions in Your Pocket Time until jam Times until jams clear

Traffic Prediction Challenge Multiple views on traffic Weather Major events Incident reports Operator ID: Nick Heading: INCIDENT Message: INCIDENT INFORMATION Cleared 1637: I-405 SB JS I-90 ACC BLK RL CCTV 1623 – WSP, FIR ON SCENE • Event store • Learning • Reasoning

From Incident Reports to Salience Operator ID: Nick Heading: INCIDENT Message: INCIDENT INFORMATION Cleared 1637: I-405 SB JS I-90 ACC BLK RL CCTV 1623 – WSP, FIR ON SCENE u e. g. , Accidents Ø How many lanes? Ø Emergency vehicles?

Data Set: Data Abstraction Lane Station Bottleneck

Building forecasting model Bayesian network. Predictive Models from Data • System-wide status & dynamics • Incident reports • Sporting events • Weather • Time of day • Day of week • Season • Holiday status Database of sensed events • Data store Max likely time until bottleneck evaporates Machine Learning • Base-level predictions

Predict Future Surprises Surprise at time t in the future: Infer surprise with model built from training set ØEvidence Ø Time of day, day of week Ø Bottlenecks, ØIncident flows and their durations reports Ø Weather Ø Games Ø Accidents

Building Models of Surprise Learning predictive model of surprise u Ø Model of user expectation and suprise Ø Models of future surprises Database of surprising events • Data store Probabiilliity P ro b a b t y ! Expectations ! Events at t Future traffic • • Major events • Weather • Time of day • Day of week • Holiday status • System-wide Human Forecaster Real World Outcome status & dynamics • Incident reports • Sporting events • Weather • Time of day • Day of week • Season • Holiday status

Building Models of Future Surprises Learning predictive model of surprise u Ø Model of user expectation and suprise Machine learning Ø Models of future surprises Database of surprising events • Data store Probabiilliity P ro b a b t y ! Expectations ! Events at t t-T Future traffic • • Major events • Weather • Time of day • Day of week • Holiday status • System-wide Human Forecaster Real World Outcome status & dynamics • Incident reports • Sporting events • Weather • Time of day • Day of week • Season • Holiday status

Adding Surprise Modeling • System-wide status & dynamics • Incident reports • Sporting events • Weather • Time of day • Day of week • Season • Holiday status Traffic forecasting service • Data store • Inference • User logs • Base-level predictions • Context-sensitive error models • Surprise forecasting models Competency annotation Surprise & surprise forecasting

Where will there be surprises in time t?

Sample Models and Exploration u Influence of Seattle sporting events

Sample Models and Exploration u Influence of weather

Sample Models and Exploration u Influence of an accident at Bottleneck 15

Anomaly at B 4 Examining Details Acc 15 . 5 surprise Influence of accidents at B 15 and B 3 on anomaly at B 4 at -low 30 minutes u Acc 3 . 5 surprise -. 25 low -. 25 high

Alerting on Device and Desktop u Specify preferences and download to device u Time- and route-specific preferences Ø Routes composed from bottlenecks Ø Multiple alert types

Models of Surprise u Definitions of surprise u Qualitative and quantitative u Abstraction about past surprises to expect classes of surprise u Wild card notion in world of futurists u Directions: Statistical models of surprise

Surprise. Cast Surprise Forecasting: 2005 -2020 Surprise Types Examples Attention & Information Top n surprise types, by class or properties Assume type occurs; give top m examples per type Recommended triage of focus of attention: additional study, info. gathering, monitoring per type Actions Recommended actions to react, contain, disrupt negative suprises and/or exploit positive surprises