The online interactive magazine of the Association for the Advancement of Artificial Intelligence

by Adam Amos-Binks

Anticipatory thinking (AT) — the deliberate and divergent consideration for relevant possible futures– is a key cognitive process in a variety of key societal contexts. The Cognitive Systems for Anticipatory Thinking brought together a multi-disciplinary group who’s methods can help augment human capacity for AT. Several key contexts emerged including autonomous vehicles, geo-political issues, and intelligence analysis. Key next steps include defining benchmark AT problems and competencies as well as how to evaluate them.

Anticipatory thinking (AT) –the deliberate and divergent consideration for relevant possible futures– is a key cognitive process in a variety of key societal contexts. From formal definitions of intelligence analysis to the strategic foresight discipline, they all rely on AT to assess the current and possible future states of the world. The COGSAT symposium convened a multi-disciplinary approach to advance methods and tools that augment AT.

Anchoring our discussions were five esteemed invited speakers. Ken Forbus kicked off COGSAT with insights on the relationship between both qualitative and analogical reasoning to AT. Gary Klein, one of the earliest thought leaders on AT, articulated the role of expertise in AT in several real world domains. Mike Van Lent described the potential of AT to augment application driven AI systems. Drew Vandeth’s history of the intelligence community’s triumphs and failures with AT underscored the importance of AT for national security. Finally, Yolanda Gil’s talk on how scientific discovery might be augmented with AT was a new domain and an inspiration to the AT researcher community.

Our accepted submissions were organized into three sessions. In the mechanisms session, Jones and Laird defined AT in the SOAR cognitive architecture using event-segmentation theory. Gilpin presented a methodology enables autonomous systems to explain possible futures that inform current decisions while Greene defined a new argument template motivated by real-world geo-political events. Rabkina et al. articulated the need for theory-of-mind for agent’s to perform AT in multi-agent settings.

Our tools and applications session featured Argenta et al.’s Scenario Explorer platform designed to augment a user’s AT with structured analytic techniques over a scalable representation. Juvina et al. showed that task-offloading tools designed to augment AT improved forecasting accuracy. The Cogent cognitive assistant from Tecuci et al. was cast as an anticipatory intelligence analysis tool while Geden et al. demonstrated the steps for using AT in risk identification. Finally, in the metacognition session Amos- Binks and Dannenhauer showed how AT exercises both meta-cognitive and cognitive cycles of the MIDCA cognitive architecture when managing risk in mission planning.

Across our three paper session, two types of AT methods emerged. The first type initiates AT in users. This modality appears to be most helpful for complex and open domain where broad expertise in one individual is difficult to draw upon and narrow expertise is readily available. A second replicates and performs AT itself. This second modality is more appropriate for autonomous systems that have a knowledge representation of the world they are operating in.

Lively discussions occurred when presenters were able to motivate their work with connections to present day AI challenges. Anecdotally, the most often discussed challenges were those pertaining to

autonomy (e.g. UAVs) and geo-political events (e.g. election results). These discussions underscored the need for the community to define competencies or challenge problems to assess the efficacy of an approach or a system’s capability. More generally – for the science of AT to take another step – the community needs to answer the question ‘What is successful AT?’ and define ways to measure it.

In addition to answering key performance questions, AT needs to continue to clearly differentiate itself from the crowded space of forecasting. There is a temptation to aggregate them together, however forecasting is only concerned with a single correct answer while AT considers and takes action knowing there are many possible futures. The NSF’s recent AI Research Institutes program is an intriguing platform where AT could be positioned as cross-cutting theme.

The symposium was organized by Adam Amos-Binks (Applied Research Associates, Inc), Rogelio Cardona-Rivera (University of Utah), Dustin Dannenahuer (Navatek, LLC), and Gene Brewer (Arizona State University). Accepted papers are available at and will be published in CEUR-WS.

Adam Amos-Binks,, Chief AI Scientist, Applied Research Associates, Inc.