AI planning

SIFT researchers were active at this year's ICAPS 2021, the International Conference on Automated Planning and Scheduling.  ICAPS 2021 was originally scheduled to happen in Guangzhou, in China, but had to be entirely virtual because of the COVID pandemic.

Creative Problem Solving Through Automated Planning and Analogy

As a creative problem solving strategy, analogical reasoning helps generalize and transfer solutions to new domains. Automated planners have been used for problem solving, but they reach impasses when their representation of the problem space lacks operators or resources to generate a plan from an initial condition to a set of goal conditions.

Marshal

Marshal is a digital personal assistant for model-based systems.  As organizations apply systems in new situations and expect new behavior, the system model may no longer characterize the environment — the model is vulnerable to drift.  Marshal reins in drifting models to ensure they best match the current environment.  Marshal combines user events, answers to clarifying questions, and experimentation to actively correct its knowledge of the system model. With up-to-date models, Marshal suggests corrections, maintains provenance, and predicts future model drift.

Maintaining Evolving Domain Models

When engineering an automated planning model, domain authors typically assume a static, unchanging ground-truth world. Unfortunately, this assumption can clash with reality, where domain changes often rapidly occur in best practices, effectors, or known conditions. In these cases, remodeling the domain causes domain experts to ensure newly captured requirements integrate well with the current model. In this work, we address this model maintenance problem in a system called Marshal. Marshal assists model maintainers by reasoning about their model as a (hidden) stochastic process.

A Hierarchical Goal-Based Formalism and Algorithm for Single-Agent Plannin

Plan generation is important in a number of agent applications, but such applications generally require elaborate domain models that include not only the definitions of the actions that an agent can perform in a given domain, but also information about the most effective ways to generate plans for the agent in that domain. Such models typically take a large amount of human effort to create.