Starting Date:
Prerequisites:
Will results be assigned to University: Yes
This project investigates how an intelligent agent can autonomously construct an interpretable model of its environment through interaction. The work will be conducted using VacuumWorld, a Python-based multi-agent simulation platform developed within the lab for studying agent behaviour in dynamic grid environments.
In VacuumWorld, agents perceive a partially observable environment and act to achieve goals (e.g. cleaning tasks) under uncertainty. While most agents rely on predefined rules or learned policies, this project explores a complementary approach: enabling agents to build explicit symbolic models of the world from their own experience.
The central focus is autoformalization – the process of converting informal or observational data into structured, formal representations such as logical rules or transition models. Recent developments in AI, particularly large language models (LLMs), suggest that such transformations may be automated to some extent. This project will examine how these ideas can be applied in a controlled simulation environment.
The student will design and implement an agent capable of:
- collecting observations of state transitions within VacuumWorld
- generating structured representations of these observations (e.g. rules or predicates)
- incrementally constructing a world model that captures environment dynamics
The resulting system will be evaluated in terms of its ability to produce accurate, interpretable, and useful models for decision-making.
Aims and Objectives
The project aims to explore the feasibility and effectiveness of learning symbolic world models in a multi-agent simulation. Specific objectives are:
- To understand the architecture and operation of VacuumWorld and its agent framework
- To design a representation for encoding observations and environment dynamics
- To develop an autoformalisation pipeline that converts observations into structured rules
- To implement an agent that constructs and updates a world model during execution
- To evaluate the usefulness of the learned model for prediction and/or control
Research Questions
- To what extent can an agent autonomously construct a formal model of its environment from interaction data?
- How effective are autoformalization techniques (including LLM-based approaches) in producing accurate and interpretable representations?
- Does access to a learned world model improve agent performance or reasoning capability?
References
[1] Sethuraman et al. Visual Question Answering based on Formal Logic. https://arxiv.org/pdf/2111.04785
[2]Towards a Common Framework for Autoformalization. Mensfelt et al., Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 40. p. 40971-40980 10 p.
[3] From Pixels to Predicates: Learning Symbolic World Models via Pretrained Vision-Language Models. Athalye et al. https://arxiv.org/pdf/2501.00296v4