Halim Djerroud

I am a researcher in theoretical computer science. My field is automated planning and epistemic reasoning: the design of algorithms that allow an autonomous agent to reason about what it knows, does not know, and can know, and to derive rational decisions in a partially observable, multi-agent setting. My work belongs to the symbolic planning community — in particular around Dynamic Epistemic Logic — and is regularly confronted with it through international competitions and conferences.

The question that structures my contributions is the arbitration between acting, observing, and delegating observation. An autonomous agent does not only have to choose what to do in the world: it must also decide when it has enough information to act, when it is preferable to produce that information through a perceptual action, and when it can rely on other sensors or other agents. This question, simple to state, opens a range of formal and algorithmic problems that structure most of my research.

A first line of work concerns the design of heuristic search algorithms for epistemic planning. The aim is to provide planners able to handle problems in which actions modify not only the world but also the knowledge state of the agents, together with heuristics that guide search efficiently without requiring admissibility in the classical sense. In this work I favour lightweight and legible heuristics over optimal ones — not out of pragmatism, but because an inadmissible yet informative heuristic often reveals the structure of a problem better than an exhaustive computation. The resulting planners are confronted with the community through the international competitions of the field.

A second line of work is more theoretical. Some classical completeness guarantees for online planning under partial observability do not transpose trivially to the multi-agent epistemic setting: the same observation may correspond to several incompatible epistemic effects, and the structural argument on which these guarantees rest no longer holds. I work on identifying the conditions under which they can be restored — not by modifying the formalism, but by identifying properties of the interaction protocol. This work brings into dialogue two communities that had long ignored each other: classical online planning, and modal-logic-based epistemic planning.

The third strand of my approach stems from an epistemological requirement. Just as a result predicted in theoretical physics is only established once it has been observed, my algorithmic and formal contributions must be confronted with the real world to demonstrate that they are more than formal objects. I therefore design and build the experimental apparatus that allows this observation: an instrumented physical environment in which autonomous agents, instantiating my planners, act in the real world. The perception infrastructure plays a dual role — a source of observations usable by the agent, and an empirical oracle against which the rationality of planned decisions can be measured a posteriori. It is this possibility of experimental refutation, rare in a community that validates primarily on synthetic domains, that gives my theoretical contributions their status as falsifiable objects.

A computational parsimony runs through this whole approach, more as a habit of mind than as a programme: prefer what is legible to what is merely optimal, engage a verification only when ambiguity demands it, observe only when no sensor already available can answer. It is less a stance than a design discipline, one that follows naturally from giving priority to decision over computation.