RAGeR

repeated analogy for goal reasoning

This project focuses on developing algorithms for chaining together analogical inferences. The Repeated Analogy for Goal Reasoning (RAGeR) algorithm has been demonstrated to do goal recognition (Rabkina et al., 2021) and to make general deductive and abductive inferences (Wilson et al., 2022).

Goal Recognition

An agent observes a Minecraft agent, and RAGeR uses analogy to infer the Minecraft agent's goal.

Goal recognition is a process of inferring an agent’s goals based on observing its actions. We use two algorithms: one based on the Analogical Theory of Mind (AToM) model (Rabkina et al., 2020), and one using RAGeR (Rabkina et al., 2021). One of the domains used to explore these algorithms is an agent in Minecraft that is gathering materials to make or obtain goods.

Deductive Inferences

RAGeR uses a sequence of analogical inferences to make deductive and abductive inferences (Wilson et al., 2022). The algorithm uses a forward chaining of inferences, searching for a specified goal. The algorithm when applied with incomplete knowledge makes assumptions about the missing knowledge when the missing knowledge is structurally consistent with the given knowledge.

References

2022

  1. Deductive Reasoning with Incomplete Knowledge via Repeated Analogies
    Jason R. Wilson, Lissangel Martinez , and Irina Rabkina
    In Proceedings of the Advances in Cognitive Systems , 2022
    Bib PDF

2021

  1. Evaluation of Goal Recognition Systems on Unreliable Data and Uninspectable Agents
    Irina Rabkina , Pavan Kantharaju , Jason R. Wilson, and 2 more authors
    Frontiers in Artificial Intelligence, 2021

2020

  1. Recognizing the Goals of Uninspectable Agents
    Irina Rabkina , Pavan Kantharaju , Mark Roberts , and 3 more authors
    In Proceedings of the Eighth Annual Conference on Advances in Cognitive Systems , 2020