Supervisor

Suitable for

Abstract

Supervised by Andrey Kravchenko andrey.kravchenko@cs.ox.ac.uk

Objective. Given (i) a user query in natural language, (ii) a world/ontology (optionally), and (iii) a catalog of heterogeneous agents, synthesise (a) a feasible set of agents and (b) a
coordination plan that achieves the user’s intent under hard constraints while optimising stated preferences.

Inputs (formal ar&facts)
• Logical intent G: Translate the natural-language query into a compact logical form -- e.g., AMR (Abstract Meaning Representation) → typed predicates -- capturing
goals, required outcomes, and any explicit constraints or preferences.
• World/ontology K: Typed predicates for skills, tasks, resources, compatibilities, regulations, and temporal/causal relations.
• Agent catalog A: For each agent: skill set, cost/risk, capacity, and (optionally) ac8on schemas with preconditions/effects.

Outputs
Team (a subset of agents) and plan (par8al order or schedule of agents and tasks) that: (1) achieves G, (2) satisfies hard constraints (skills, capacity, mutual exclusions, resource,
temporal), and (3) is op8mal under a declared objective (e.g. cost, makespan, risk) and preference model. In ASP this is naturally expressed via weak constraints/preferences.

Expected background of students and CS techniques that will be applied
1. Expected background of a student:
       1. strong Python programming;
       2. good CS/AI fundamentals (algorithms, search, optimization, planning);
       3. basic logic / symbolic reasoning (predicates, constraints);
       4. some NLP / LLM familiarity (useful for parsing natural-language tasks);
       5. ability to run experiments and write clear technical documentation;
       6. nice to have: ontologies/knowledge graphs, ASP/CP/SAT/SMT, multi-agent systems, scheduling.
2. CS techniques that might be applied:
      1. NLP / semantic parsing: convert user request into structured goals and constraints;
      2. knowledge representation: predicates, ontologies, task/agent/resource graphs;
      3. planning & scheduling: task decomposition, dependency handling, temporal planning;
      4. constraint solving / optimisation: CP, ASP, SAT/SMT, MILP;
      5. preference handling: soft constraints, weighted objectives, multi-objective optimisation;
      6. agent/team selection: matching, coverage, assignment, heuristics;
      7. hybrid neuro-symbolic approach: LLM for parsing + symbolic solver for planning/validation