Differences

This shows you the differences between two versions of the page.

--- 2023-24:winter [2023/12/28 19:41] – eruppert
+++ 2023-24:winter [2024/01/05 15:32] (current) – eruppert
@@ Line 560: / Line 560: @@
 Please send your c.v., transcript, and Statement of Interest in the project to the project supervisor
+==== Investigating Reasoning about Action and Change in Pretrained Language Models ====
+**Course:**  EECS4080
+**Supervisor:**  Yves Lesperance
+**Supervisor's email address:**   lesperan@yorku.ca
+**Project Description:**
+Reasoning about action (RAC), including generating plans to achieve goals, is a key capability for autonomous agents.  Mainstream techniques for RAC and automated planning (based on heuristic search), are very effective, but they rely on a human modeler specifying the dynamic domain and queries/goals formally. Some recent research has investigated whether pretrained language models (LM) can effectively reason about action and change while avoiding the need to formally specify the domain.  For instance, He et al. (ACL 2023) has studied the performance of some LMs on fundamental RAC tasks such as Projection, Executability, Plan Verification, and Goal Recognition.  The LMs were first fine-tuned/pretrained on Blocks World domains and task instances (with a STRIPS semantics) and then tested on new instances.  The LMs performed rather well on similar instances, but generalized poorly to tasks involving longer action sequences or more domain objects.  In this project, the student will use the datasets generated in this work to experiment with newer language models and various fine-tuning methods to see if generalization can be improved.
+**Required skills or prerequisites:**
+EECS3401, Python programming skills.
+**Recommended skills or prerequisites:**
+Some previous exposure to machine learning, automated planning, first-order logic.
+**Instructions:**
+Send CV and unofficial transcript to project supervisor.