COGSAT 2021

Symposium Aims

Join our Slack channel here!

Register here!

Anticipatory thinking – the deliberate, divergent consideration of relevant possible futures-- is a key cognitive process for risk management. The cognitive systems community has taken steps towards incorporating risk management into solutions and designing scenarios where risk is present. These initial efforts have demonstrated the value to AI systems incorporating risk management into decision-making.

These cognitive system efforts to manage risk have largely been in isolation without real-world deployments. In contrast, data-driven AI systems are being deployed in the real-world but without the tools for risk management. The lack of tools results in inconsistent regulatory and legal policies that limits the widespread adoption of AI systems, increases the risk to the public, and sows mistrust of autonomous systems.

This year’s COGSAT brings together cognitive systems and statistical learning communities to develop risk management capability for 3^rd wave – DARPA defines this as context-adaptation -- autonomous systems. Autonomous agents with 3rd wave risk management capability will make decisions similar to an insurance agent assessing premiums; identify perils for a context (life, auto, home) and when the risk of these perils change. However, 3rd wave agents would constantly be re-assessing risk and mitigating them with a variety of actions.

Specifically, we seek the community's input to achieve the following goals:

1. Refine the perception and cognition challenges to spur research in AT and risk management

2. Identify benchmark domains, measures, attributes, and metrics for AT

3. Develop 3rd wave autonomy capabilities based on AT

Picking up where COGSAT 2019 left off, COGSAT 2021 will introduce two challenges in perception and cognition with an example domain in each.

In the perception challenge, an autonomous vehicle requires pre-hoc assessment of errors and risk management. An example research question is how to assess, test, and evaluate a perception system in a self-driving car for the ability to handle out-of-sample and never-before-seen images. Perception in self-driving cars remains unsolved; there are many well-documented perception errors where anticipatory thinking, in the form of self-explanation, can help mitigate error and failure states.

The cognition challenge examines the disconnect between an agent’s action model and changes in risk exposure. We use Dungeon Crawl Stone Soup, a character-development game, as an example domain due to its catastrophic failures (permanent death) and many opportunities to manage this changing risk. The research question is how to identify when risk changes and mitigate exposure.

For more details on the Scope and Recommended Readings of our community, please visit the Home page.

We expect this symposium to draw on prior work from existing AI fields including but not limited to: mixed-initiative planning, goal reasoning, case based reasoning, analogical reasoning, computational narrative, intentionality, theory-of-mind, cognitive architectures, games, intelligent tutoring systems. In particular, we are excited to provide a convergence point between the cognitive systems and prospective cognition communities. This focus on collaboration and strong application contexts, offers several ways to engage. Our symposium will include:

• Lighting talks (10 - minutes) for abstracts accepted for presentation

• Invited talks (45-60 minutes) for guest speakers

• Panel discussions (60 minutes) for perspectives and thoughts to seed bull sessions

• Bull sessions (90 minutes) bid for time to spur collaboration opportunities

Participate

A full call for participation is available here, below are the important parts:

We invited contributions in a variety of forms on topics within the scope of this call:

• Abstracts (3 pages + 1 for references) describing an approach, idea, theory for 3rd wave autonomy.

• Accepted abstracts will be invited to submit an expanded post symposium for archival purposes published via CEUR

• Review guidelines are here (we recommend answering each question explicitly, example submission here)

• Paper submissions should follow AAAI style guidelines.

Papers will be uploaded through EasyChair and will be guaranteed at least two reviews. Further, papers should be format for blind review with all identifying author information removed. You are welcome to upload your paper on your own personal website or ArXiv with identifying information.

Attendee information can be found at the main AAAI site here.

Important Dates (updated July 23)

All deadlines were specified as Anywhere on Earth time unless otherwise noted.

• Submissions due: Friday September 3 EXTENDED to Friday October 1

• Notifications to authors: Friday September 17 EXTENDED to Friday October 22

• Registration deadline: Friday October 15 EXTENDED to Friday November October 29

• Final papers were due to organizers: Friday October 1 EXTENDED to Friday December 3

Organizing Committee