Artificial Intelligence

How does a system move from recognizing patterns to deciding what to do next? Our work in artificial intelligence asks how agents can represent goals, reason about consequences, and adapt their behavior as environments change around them.

Research topics

  • Reasoning and planning
  • Intelligent agents
  • Multi-agent systems
  • Knowledge representation
  • Decision-making
  • Reinforcement learning
  • General-purpose intelligent systems
Fig. 01Schematic — not experimental data
s₀goald=0d=1d=2d=3d=4
Fig. 01Schematic. A search tree over possible actions; the highlighted branch marks one plan selected under a learned value estimate.

Why it matters

Why it matters

Most deployed machine learning still maps inputs to outputs in a single step. Many of the problems that matter — scientific investigation, engineering design, coordinating with people — instead require sequences of decisions in which each step changes what is possible next. Getting those sequences right depends on capabilities that pattern recognition alone does not guarantee: search, abstraction, credit assignment, and the ability to notice when a plan has stopped working.

Agents are also where questions about reliability become concrete. A system that acts in the world accumulates the consequences of its own errors. Understanding how agents plan, and how that planning fails, is a prerequisite for building systems that can be given meaningful responsibility.

Key research questions

Key research questions

  1. Q1

    When does explicit search or planning outperform a learned policy, and can a single system decide which to use?

  2. Q2

    What internal representations allow an agent to transfer a skill learned in one environment to a structurally similar one?

  3. Q3

    How should an agent represent uncertainty about its own goals or about the intentions of other agents?

  4. Q4

    What are the right abstractions for long-horizon tasks, and can they be discovered rather than designed?

  5. Q5

    How can we evaluate an agent's competence in open-ended environments where success is not a single number?

Primary subfields

Primary subfields

Reasoning and planning
Search, deliberation, and the use of world models to evaluate consequences before acting.
Reinforcement learning
Learning behavior from interaction and delayed feedback, including exploration and credit assignment over long horizons.
Multi-agent systems
Cooperation, competition, and communication among agents with partially shared information.
Knowledge representation
Structured, compositional representations that support inference beyond the training distribution.
Decision-making under uncertainty
Acting well when outcomes, preferences, or the environment itself are only partially known.

Representative methodologies

Representative methodologies

  • M1

    Model-based learning

    Learning predictive models of an environment and planning against them.

  • M2

    Procedurally generated environments

    Families of tasks with controlled structure, used to separate memorization from generalization.

  • M3

    Hybrid neural–symbolic systems

    Combining learned representations with explicit structure where structure is known to help.

  • M4

    Behavioral evaluation

    Measuring competence through diverse, held-out scenarios rather than single benchmark scores.

Relationships to other domains

Relationships to other domains

Working on questions in Artificial Intelligence?

We welcome conversations with researchers and organizations working on related questions — whether to compare notes, discuss a collaboration, or share work in progress.