Halim Djerroud

My research lies at the intersection of automated planning and epistemic reasoning, with a particular focus on autonomous decision-making under partial observability and in multi-agent settings. The question that structures much of my work is the arbitration between acting, observing, and delegating observation: determining when an agent should act on its current knowledge, acquire additional information, or rely on information produced by another agent or sensor. My work is primarily grounded in symbolic planning and Dynamic Epistemic Logic.

On the algorithmic side, I investigate heuristic search for both classical and epistemic planning. S-Planner provides a classical planning framework, while DEPTH addresses epistemic planning problems in which actions may affect not only the state of the world but also the knowledge of the agents involved. A recurring concern in this work is the design of lightweight and informative heuristics that expose useful problem structure without making optimality or admissibility an objective in itself. DEPTH received the Runner-Up Award at the First International Epistemic Planning Competition (IEPC 2026).

A second line of research concerns the theoretical foundations of online epistemic planning. Classical completeness arguments for planning under partial observability do not transfer directly to multi-agent epistemic settings, where a single observation may induce several incompatible epistemic effects. I study the structural conditions under which such guarantees can be recovered, in particular through properties of the interaction and observation protocols. This work connects questions traditionally studied in online planning with those arising from modal-logic-based epistemic planning.

The multi-agent dimension of these questions is explored through gAgent, a framework I develop for modelling and experimenting with interacting autonomous agents. It provides a setting in which communication, information production and information transfer can be studied together with planning decisions, and in which delegated observation can be treated as an explicit component of agent behaviour rather than solely as part of the domain model.

I also consider experimental confrontation with physical systems an important part of this research programme. Formal and algorithmic results become more informative when assumptions about observation, uncertainty and interaction are exposed to the constraints of a real environment. UbiMap provides an instrumented multi-robot environment and an external perception infrastructure that can serve both as a source of observations and as an empirical reference. Ratbot provides an open robotic platform on which planning and autonomous decision mechanisms can be instantiated in the physical world. Together, these systems complement evaluation on formal domains and benchmarks with experiments in which perception and action have concrete physical consequences.

A principle of computational parsimony runs through these different lines of work: favour representations and heuristics that remain intelligible, introduce verification when ambiguity requires it, and acquire new information only when the information already available is insufficient for decision. The common objective is not computation for its own sake, but the design of agents able to reason explicitly about information and use it economically in order to decide.