Avoiding Surprise: Augmenting Anticipatory Thinking with Scenario Explorer Chris Argenta, PhD; Abigail Browning, PhD; Adam Amos-Binks, PhD, and Matthew Lyle Applied Research Associates, Inc. Raleigh NC cargenta@ara.com Abstract In this research, we do not attempt to divine things about After being unpleasantly surprised in a high-value and high- that future that cannot otherwise be known to the people in- risk situation, we often hear phrases like “we had all the dots, volved. Instead, we focus on developing techniques to help but we just didn’t connect them correctly.” This is an exam- avoid surprises that result from incomplete prospective cog- ple of a failure in anticipatory thinking (AT), the people in- nition. After being surprised, how often do we hear state- volved apparently knew all of the facts they needed but did not collaboratively reason over them rigorously enough. This ments like these? situation differs from when possibilities are considered but “We had all the dots, but we just didn’t determined to be too unlikely to act upon, the first is true sur- connect them correctly.” prise while the latter is risk management. “Some people expected X, others expected Y, In this paper, we examine some of the cognitive founda- but no one expected X and Y.” tions of surprise with an eye towards identifying analytic techniques that augment the support experts in systematic and “We did not think X was possible, but after Y happened collaborative anticipatory thinking. We describe Scenario we should have reconsidered at X.” Explorer, our prototype cloud-based collaborative imagina- These are likely cases where rigorous collaborative pro- tion support platform. We present a set of cognitive-informed spective cognition may have helped participants to avoid be- analytics techniques within Scenario Explorer augment the working memory, manage attention through systematic rea- ing surprised. Our platform, Scenario Explorer, integrates soning, and communicate prospective futures between dis- techniques that attempt to address this class of surprises and tributed participants. shift the situation from one of complete surprise to one of managed risk though Anticipatory Thinking (AT). Introduction to Avoiding Surprise Motivation: Managing Risk by Avoiding Surprise Some surprises are pleasant. However, in high-value and The 2019 National Intelligence Strategy (Coats 2019) lists high-risk endeavors surprises tend to be feared rather than “Anticipatory Intelligence” as the second of seven key mis- enjoyed. Throughout history, people concerned with the sion objectives. It lists foresight, forecasting, and warning downside of being surprised have employed a myriad tech- essential elements of Anticipatory Intelligence. While data niques to divine the future, with wildly varying efficacies analytics is clearly critical for the Intelligence Community (see Pickover 2001 for an interesting survey of such tech- to stay ahead of world events, many of future situations of niques). In fact, organizations spend fortunes trying to stay interest have not been previously observed and are therefore ahead of the trends and predict future events in order to bet- difficult to predict using only analytics. This places empha- ter position themselves and avoid being caught unprepared. ses on the subject matter experts to interpret the data for sit- Today, predictive analytics is a huge field dedicated to the uational awareness, imagine the future trajectories that may computationally divining the future. Despite all of these ef- evolve, and identify key indicators that might identify which forts, we continue to be surprised and caught unprepared, of those trajectories are occurring in the future. Anticipatory even in high-value and high-risk situations where such sur- Intelligence attempts to do this well in advanced of the ac- prises can be catastrophic. tual situation to provide time to understand and prepare. Copyright © 2020 Applied Research Associates, Inc. All rights reserved. Use permitted under Creative Commons License Attribution 4.0 Interna- tional (CC BY 4.0). Risk management is the process of identifying potential ing mechanism triggers the emotional response we call sur- futures, estimating the likelihood and impact of these fu- prise. Surprise highlights the discrepancy, and brings ones tures, and designing mitigations to manipulate the likelihood attention to bear on the process of determining what should and/or impact of trajectories. We generally try to encourage be done to bring reality and ones frame into harmony again. futures we rate positively and avoid those we rate nega- Related research (see survey in Reisenzein 2019) at- tively. While we often cannot mitigate all risks, the process tempts to qualify and quantify surprise. For example, as- of risk management seeks to assess trade-offs and optimize sessing the magnitude of the surprise with respect to this dif- our mitigation efforts. However, we can only perform effec- ference, or creating experimental conditions to elicit sur- tive risk management when we can enumerate the potential prise. While, none of this research appears to be focus di- future trajectories in some form. We can accept risks, but rectly on avoiding surprise, the concepts defined in this ap- this is a choice or trade-off. We can be surprised when our proach appear to lend themselves to this application. estimates of likelihood or impact prove incorrect, but these If we accept both of these cognitive models, than we can are a matter of quantifying uncertainty. However, when the identify two key places for which intervention might reduce realized future is significantly different from any of the po- the occurrence of surprise and/or minimize its magnitude: tential futures that we have anticipated, we are surprised, we • Frame Curation - If, for a situation of importance, we lose situational awareness, and we have to reassess our risks. can improve one’s frames/schemas to support a wider Unfortunately, after being surprised in this way, we have range of feasible situations, then we might expect there to less time, more pressure, and frequently fewer options for be fewer discrepancies. Similarly, when there are discrep- mitigation or planning. ancies, we might want alternative frames to which we can We designed Scenario Explorer to include techniques that more readily switch. systematically elicit and aggregate feasible (sometimes un- • Discrepancy Awareness - If we can instill into relevant likely) futures from multiple participates, automate the pro- frames a sensitivity to key features for discrepancy mon- cess of systematically estimating risk, and analyzing the re- itoring, then we might expect to catch frame-changing differences earlier. In the IC, Indicators and Warnings are sulting trajectories to extract key indicators and warnings to methods for characterizing these features or events. support situational awareness. Since we natively develop frames through experience and knowledge, we can infer that people with more varied expe- Some Cognitive Foundations of Surprise riences and knowledge should be surprised less often. Pos- The Klein model of Sensemaking (Klein 2007) proposes sessing a set of related and diverse pre-existing frames might that the mind creates and uses data structures called frames further reduce the delay and cognitive burden of formulating (sometimes also called schemas) that allow us to store facts a new frame while in the midst of surprise. As shown in Fig- in a context. These frames represent our cognitive model of ure 1, surprises result when our frame anticipates one thing a situation based on our previous experience and by we are presented with another. The objective is to culti- knowledge. Frames embed knowledge about how to under- vate variations for important frames so we can easily recog- stand things we have observed, how we interpret some am- nize when a frame change is needed and have alternative biguous knowledge, and what we expect to see in the future. pre-considered frames readily available. Because of working memory limitations, we generally focus on a single frame. When our frame matches the real world situation, these mechanisms are effective. However, when our current frame fails to match the real world situation we tend to misunderstand, misinterpret, and mispredict. Under the Sensemaking model, when we recognize that our current frame does not fit the situation, we can decide to abandon our current frame and seek another, or modify our current frame to accommodate the new situation. The Cognitive-Evolutionary Model of Surprise (Meyer1997) relates to the sensemaking in that is posits an innate and unconscious cognitive mechanism that monitors the alignment of our current cognitive frame and our obser- vations of the world. When the observations match the ex- pectations everything works smoothly just as in the Klein model. However, when the real world observations conflict with the expectations derived from the frame, this monitor- Figure 1. Observations trigger our frames, but surprises result when our frames fail to fit the observations. One challenge is in effectively sharing those frames and • Enabling multiple analysts to work together to converge composing elements of different frames. Storytelling is one on a common model of the features, value systems, and transfer mechanism for situations that are not easily experi- timeframes that are relevant for a given project or topic. enced directly. Use of a branching narrative allows us to • Eliciting the known, expected, or previously imagined fu- consider many potential trajectories and how they differ. ture events and scenarios while expressing them consist- ently with respect to the common model. • Putting the analyst in a mind-set to imagine feasible fu- The Role of Anticipatory Thinking ture events that have significant effects on the features in in Avoiding Surprise the common model. • Combining the elicited scenarios, automatically compos- Anticipatory Thinking (AT) is intentionally divergent think- ing the events in scenarios (potentially created by differ- ing that enables a person to better foresee future events (and ent users) to generate novel but sensible scenarios. combinations of events) and their cascade of consequences. • Intelligently querying the combined scenarios to discover If we could always correctly predict the sequence of events key events and potential leading indicators that can deter- that will occur, there would be little need to consider any mine which scenarios are likely to be occurring. other scenarios. However, we often view the evolution of • Exploring the effects of mitigations and manipulations to the situation with limited and noisy information, so despite assess the sensitivity and uncertainty in trajectories, and our best predictions, we are frequently surprised. AT is sim- the value of triggering potential interventions. ilar to prediction and forecasting because they all attempt to The primary goal of an imagination support is to automate correctly identify the scenario that is evolving. However, as and augment the ability of analysts to conduct these kinds of shown in Figure 2, prediction and forecasting tend to prior- techniques interactively, reliably, and at scale. The Scenario itize Precision (being close to the correct answer and avoid- Explorer platform is being designed and prototyped to bal- ing false alarms) while AT prioritizes Recall (ensuring the ance the cognitive load of the analyst with the computational correct scenario in the set of answers). These different pri- power of modern computers. orities lead for different analytic techniques. Imagination Support with Scenario Explorer One thing that forecasting and AT have in common is the understanding that the future can be significant altered by the sequence of events and states that occur. Conditional forecasts are predictions of the future assuming some given Figure 2. The trade-off of Precision and Recall defines a space of condition is true. For example, a general forecast question future-oriented analytics. On the left, we strive to give a single might be “what will the average price of an electric car be in close answer to achieve high precision. While on the right, we 2030?” while a conditional forecast might be “what will the strive to ensure the single correct answer is included. average price of an electric car be in 2030, assuming that Tesla releases the Model 3 as advertised?” The additional Foresight helps prevent us to finding ourselves in a situa- qualifier limits the scope of the forecast question because we tion we had not previously considered or imagined. So, if no longer have to consider alternatives possibilities for the we can imagine the future, why are we still so often sur- Tesla Model 3 release. Conditions qualify a forecast by list- prised by it? Often the answer to this question lies in the ing the situations under which it is believed to be correct and complexity of: talking through the potentially relevant improving the accuracy of a forecast can sometimes be de- things that you (or your group) know; rigorously and sys- pendent on identifying the conditions under which we can tematically developing a set of scenarios; and identifying expect it to be valid (usually the assumptions under which it which scenarios are of interest with respect to various con- was analyzed and computed). One can understand the im- cerns. Human working memory limitations and time con- pact of a conditioning by comparing the forecasted data with straints reduce us to thinking about a small sample from the the qualification and without, or against forecasts with other range of feasible scenario. This sampling is often biased to- conditions. In previous research, we showed that distinct wards those things that we have already imagined, if not modalities in crowd-sourced forecast responses can often be come to expect. explained by differences in the conditions assumed by the ARA’s Scenario Explorer platform integrates several forecasters. novel structured analytic techniques that attempt to help an- In AT, we work to develop a divergent set of possible fu- alysts more rigorously, efficiently, and creatively explore a tures, so identifying interesting conditions that would cause wide range of feasible future scenarios by: us to adjust our forecasts of future states is a primary con- cern. As shown in Figure 3, a “conditioning event” is a sit- ways to create previously unconsidered scenarios. With AT uation in which there are multiple possible and mutually ex- we consider feasible sequences of events and interactations. clusive outcomes, each outcomes resulting in a different ef- In Scenario Explorer, we represent possible futures as a fect on the values that would be forecast for a future state. tree strcture. There are nodes are States that are hold the val- Grouping related outcomes under a common semantic event ues for each feature. States are anchored in both their time allows us to systematically evaluate each possible outcome. and their position in the tree of possible futures. The root of By quantifying the effects that each outcome has on the fea- the tree represents our current time (i.e., “Now”) and in- tures allows us to trace changes the state of the world (within cludes the current values for each feature. Each trajectory the confines of this project) depending on which outcomes can be viewed as a multi-variate time series covering includ- conditions it. ing the changes of feature over time. Projectors are algorithms that forecast the future value of their assigned feature based on its previous states. An exam- ple of a projector could be a function that adds compounding interest and this projector could be assigned to a feature like the amount of money in ones savings account at the bank – even with no intervening events, the amount of money in the account will increase over time. If no projector is assigned then the feature value is propagates unchanged until a con- ditioning event outcome changes it. Assessing Risk requires both the likelihood and the im- pact or value of a State. We define Values/Impacts as a score (0.0-1.0 with an associated color scale) and we train a model to score each State in the tree. This allows display and clus- Figure 3. A conditioning event changes the context in which fu- tering based on the applied value scale. ture events occur in a predictable way. A conditioning event can Conditioning Events introduce branching to the tree of fu- only occur when its pre-requisites are satisfied. It has (at least 1) tures. A scenario (or trajectory through the tree) in which possible outcomes, each that has a set of (at least 1) effects. The outcomes branch the futures tree, and their effects modify the val- Outcome #1 occurs is different from the one where Outcome ues of features between the previous and next state. #2 occurs. We assign likelihoods to each outcome (and a null outcome), resulting in each trajectory have some likeli- For example, let a feature (A) represents the number of hood. Conditioning Events can be composed in multiple electric cars sold per day in some area. Today, we may sell ways in the tree to produce many possible futures; however 20 cars per day and we are interested in what future sales must fit the Conditioning Event’s pre-requisite constraints. might look like. We identify that announcement expected Mitigations are actions that can manipulate the likeli- next week about federal subsidies for electric cars would hoods (e.g., make one outcome more or less likely) and ef- likely have a significant effect on that number in the future. fects (e.g., reduce a feature value change) of conditioning We can imagine an announcement that increases incentives events in the tree or the value/impact of states (e.g., make could cause the sales numbers to increase. An alternative some characteristics of a state more or less negative). Unlike outcome could be removing incentives which could have the conditioning events, each mitigation is either on or off. effect of the sales number dropping. Consider two represen- tations: Storytelling through Futures Building • In forecasting, we might write two conditional forecasts The primary goal of Futures Building is to elicit interest- – one would say “what would the sales of electric cars be ing and feasible Conditioning Events from the analyst and next week, if the announcements increased incentives?” display them. Futures Building accomplishes this by ena- and the other would end with “if the announcement re- bling an analyst to express a scenario that they believe to be moved incentives?” The focus is on a feature value given relevant to the project. This technique starts with a tree con- a specific context. taining only the root Now node. The analyst builds a set of • In anticipatory thinking, we qualifying when or how the scenarios by sequentially adding conditioning events that announcement event it might occur, enumerating the pos- logically fit together as a form of narrative (albeit in tree sible outcomes, and estimating their effects on the fea- form). While the analyst may be focused on expressing a tures relative to their previous values. Here, the focus is on defining potential effects in a general context. single scenario (i.e., a single trajectory through a tree of pos- sible futures) that is relevant to them, the system is automat- These may seem like minor differences, but the condi- ically populates the tree with all possible compositions of tioning event representation allows us to take conditioning the events elicited. events created in one context and compose them in new Cognitive Basis variation or divergence desired. If the conditioning events Futures Building elicits branching narratives inde- required to reach the goal already exist in the project, Sce- pendently from multiple participants decreasing anchoring nario Explorer detects them and fills in the trajectory. and priming effects that might occur during brainstorming. Value Added When to Use It It is sometimes difficult to diverge from the events we ex- Futures Building is useful when the analyst has a set of pect. Extreme States provides a method of shifting ones per- conditioning events that they wish to enter into the system spective from looking forward at a wide-open expanse of that share common narrative thread. Ideally, each view in- possibilities to looking back and explaining how the situa- stance would stand on its own as a short story of what might tion got to that point. The trajectory directly elicited is not happen in the future. This is the default elicitation view for the primary objective since it likely highly unlikely, how- the Imagination Support platform. ever the individual conditioning events introduced can also Value Added be used to enrich other scenarios and introduce trajectories When analysts express their knowledge as conditioning that are less extreme and more likely. events in the Futures Building view, they are sharing their The Method knowledge with the team and the system using a common Scenario Explorer treats each instance of Extreme States project model. A Futures Building session can be performed as a separate View. An Extreme States view will likely be alone or as a team. When/if the entered conditioning events relatively small, since once a trajectory between Now and appear on other views they can be traced back to their origin the Extreme State is found, the process is complete. view – a Futures Building view provide a way to group re- • Step 1. Creating an Extreme State View lated conditioning events to more effectively explain their • Step 2. Specifying the Extreme State meaning by offering context. • Step 3. Checking for an Existing Trajectory The Method • Step 4. Adding Conditioning Events Scenario Explorer treats each instance of Futures Build- ing as a separate View. Users can collaborate on a shared Smart Queries Extract Warnings and Indicators Futures Building scenario or refer back to it to understand Smart Query is a technique for extracting knowledge the intention of a conditioning event that might come up in from Scenario Explorer rather than eliciting knowledge another context within the system. Behind the scenes, the from analysts to put into the system. Smart Queries allow conditioning events being elicited are automatically being one ask questions about a large number of trajectories in the applied at the project level. So, instead of one large Futures tree in an intuitive manner. It provide results based on clus- Building View, we recommend that users create many ters of terminal states (based on similarity to query features smaller/simpler Views around specific domain-relevant nar- or values/impact scores) and identifies and scores sensitivity rative topics. and selectivity of conditioning event sequences that best dis- • Step 1. Create a New Futures Building View criminate these clusters. • Step 2. Add New Conditioning Events For example, if we perform a Smart Query on a feature • Step 3. Integrate the Conditioning Event into a Tree that represents the average price of electric cars, all futures could cluster into ranges of the cars being generally expen- Using Extreme States to Change Your Perspective sive, mid-range, or cheap. We might find that conditioning The Extreme States technique derives from the concept of events and outcomes such as “major tax incentives given” a Pre-mortem analysis (Klein 2007). Rather than starting would exist on the trajectories for mid-range and cheap, but with a blank slate and working forward, Extreme States not expensive. starts with a goal State and attempts to elicit conditioning Cognitive Basis events that bring trajectories closer to this goal state. This is Identifying patterns that lead to futures helps analysts be- intended to aid the analyst in shifting their perspective to come sensitive to key events that indicate a class of futures. imagining how the situation might have gotten from Now to When to Use It the Extreme State. A Smart Query should be used when the analyst wishes Cognitive Basis to determine which conditioning events play a critical role Changing your perspective helps trigger frames that may in differentiating possible futures with respect to a specific be less accessible when in a different mindset. set of features. It allows the analyst to gain new insight ei- When to Use It ther by seeing how conditioning events and outcomes (re- Extreme States should be used after analysts have ex- gardless of their source) influence futures. hausted the conditioning events they previously expected but the tree of futures still does not contain the degree of Value Added overlay on the project in which they can modify specific fea- Smart Queries extract short stories from large data sets by ture values and conditioning event outcome effects at will finding the most salient conditioning events and outcomes and without changing other views. for understanding how a feature might evolve. The critical Cognitive Basis conditioning events and outcomes can be used as leading in- Understanding sensitivities in the futures considered dicators to determine which cluster an evolving scenario is helps establish confidence bounds and identify assumptions likely to be in. that may invalidate our analyses. The Method When to Use It Scenario Explorer treats each instance of Smart Queries A What If Analysis should be used when an analyst has as a separate view, since they can be computationally expen- questions about specific features and how sensitive things sive to produce; the query results are stored and updated on are to those values. For example, they can change the Now demand. state and see the effects ripple through the tree. • Step 1. Creating a Smart Query View Value Added • Step 2. Specifying the Query Features or Value/Impacts A What If Analysis allows an analyst to explore nuances • Step 3. Execute the Query of a specific feature or view of the project. These overlays do not modify or feed back into the data. Risk Mitigation Analyses The Method Risk Analyses allows analysts to compare the effects of Scenario Explorer allows What If Analysis Views to performing (or not performing) specific mitigation actions modify feature values in states generated from any other in response to triggering events within the trajectory tree. view. It does this by cloning the trajectory data and only re- Comparing a customizable plot the trajectory risk scores calculating the portions of trajectories that are changed. side by side allows the analyst to visualize the risk reduction. • Step 1. Create a What If Analysis Overlay Cognitive Basis • Step 2. Edit the values of features in any State Risk Analyses allows users to experience the effect of po- • Step 3. Update the tree with the new consequences tential mitigation actions they imagine thus attaching addi- • Step 4. Compare the original and overlay trees tional action-oriented knowledge to relevant frame. When to Use It A Risk Analyses is used when the analyst wishes to un- Future Work derstand the effect that a set of specific actions might have While many of the capabilities outlined in this paper exist on the future. This can used to determine which mitigations today, many are also in a state of partial implementation, and are most effective and what risks might be immune to them. others only mostly design. Our plan is to complete the out- Value Added lined capabilities and evaluate our platform with analysts A Risk Analyses automates the process of applying a set and subject matter experts. of actions in the proper context for a complex set of future Additionally, we have planned to continue the develop- trajectories. Even with simple actions, this often taxes work- ment of Scenario Explorer to include automated handling of ing memory and is highly error prone when done manually. historic and streaming data to enable updating of our trajec- The Method tories over time and improve projector performance. Scenario Explorer performs Mitigation Analyses by com- Finally, we believe that we have just opened the door to paring the likelihoods and impacts of potential states both vast opportunities for additional AT analytic techniques. As with and without a selection of mitigation actions enabled. we continue to develop analytics we have designed, we wish • Step 1. Create Mitigation Analysis View in a project. to open our platform up for other researchers in the AT com- munity to contribute to the tools and execute experiments • Step 2. Define a set of Mitigations/Interventions that leverage the quantification capabilities of platform. • Step 3. Toggle the mitigations that are enabled Please feel free to contact us to discuss collaborations. • Step 4. Compare the plots and analytic summaries What If Analyses for Understanding Sensitivities Conclusions What If Analysis is used to explore the sensitivities of the Today, we are free to imagine the future in highly uncon- scenarios represented in the tree. A What If Analysis starts strained ways. However, this freedom comes with some with an existing view and allows the analyst to create an costs, which include not having systems that support, ex- tend, and collaborate your imagination. As a result, our prefactual reasoning is limited and we are too frequently left surprised. This is problematic in high-value and high-risk situations. This is more made more frustrating when we have the knowledge we needed to avoid the surprise but simply failed to apply enough rigor to our analyses. By understanding cognitive models underlying sense- making and the experience of surprise, we can design tools that help us avoid surprise and better manage risks. Scenario Explorer is one such tool, and we have outlined five struc- tured analytic techniques that we have developed to improve anticipatory thinking and to avoid being surprised. Acknowledgement The authors wish to acknowledge that this research has been funded by the Laboratory for Analytic Sciences (LAS). We would like to thank LAS and our government, academic, and industry partners for their support and cooperation in creat- ing and nurturing a research community around Anticipa- tory Thinking and for their support of this research effort. References Coats, D. 2019. National Intelligence Strategy of the United States of America. Office of the Director of National Intelligence, DC. Klein, G., Phillips, J. K., Rall, E., & Peluso, D. A. 2007. A Data/Frame Theory of Sensemaking Klein, G. 2007. Performing a project PreMortem. Harvard Busi- ness Review, September, 18–19. Meyer, W.-U., Reisenzein, R., & Sch€utzwohl, A. 1997. Towards a process analysis of emotions: The case of surprise. Motivation and Emotion, 21, 251–274. Pickover, C. 2001. Dreaming the Future: The Fantastic Story of Prediction, Prometheus Books, New York Reisenzein, R., Horstmann, G., Schützwohl A. 2019. The Cogni- tive-Evolutionary Model of Surprise: A Review of the Evidence. Topics in Cognitive Science. 2019 Jan;11(1):50-74. doi: 10.1111/tops.12292. Epub 2017 Sep 23.