Studying how agents learn hidden rules and respond to change.
If the rules of an environment change, can an AI agent notice, revise its understanding, and adapt its actions?
A controlled simulation that records predictions, actions, and actual outcomes, supported by model-training experiments and comparisons with simpler methods.
A working testbed and a clearer distinction between predicting correctly, choosing a useful action, and recovering from false assumptions.
I directed and built the research independently. It began with a broader interest in how agents form a model of the world: what exists, how things relate, and what actions are possible. Simulagent made one part of that question testable by focusing on prediction and adaptation to environmental changes.
Build worlds with explicit rules, so the simulation determines what happened independently of the agent’s explanation.
Pair a changed environment with a working control. An agent should respond to the change while still making progress in the control.
Compare expected and actual outcomes, then inspect whether a mistake came from prediction, action choice, or recovery.
Test alternatives, audit the evaluation itself, and narrow the claim when the evidence doesn’t support the original idea.
Predict
Compare its account of the world with the simulator’s actual transitions.
Act
A correct written prediction doesn’t necessarily produce a useful action.
Adapt
Inspect whether feedback changes the next decision or simply repeats the same mistake.
I compared working scenarios with ones where progress was blocked, such as a hatch relocking or machinery losing power.
These paired situations helped distinguish adaptation from blanket caution. Avoiding every risky action wouldn’t count as understanding the environment.
A separate follow-on, Executable Ontology Lab, explored how agents might construct and revise descriptions of a world that could be used for prediction and planning.
I paused the broader investigation when progress required additional datasets, independent benchmarks, and external validation. Adapting to entirely novel ontologies remains an open question.
This project taught me to design the evaluation alongside the system. Understanding where an agent fails is often more useful than knowing whether it eventually reached the goal.