Reinforcement Learning & Agents Center

The Reinforcement Learning & Agents Center teaches sequential decision-making theory and autonomous agent architecture spanning deep RL, model-based world models, multi-agent systems, and LLM-based agents. Students formulate scientific exploration, experiment design, and adaptive control problems as reinforcement learning problems and train agents to solve them, progressing from single-agent settings to multi-agent collaboration. Through this progression, graduates reach the point where they can design and implement autonomous systems that make their own decisions in uncertain, changing environments.

⚙️ GeoAcademy System Administration College 🖥️ AI & Computing Sciences College 🔬 Basic Sciences College ⚡ Intelligent Geophysical Exploration College 🛢️ Resource & Energy Engineering College 🌍 Applied Geoscience Solutions College 💼 Economics, Policy & Strategy for the Future College 🎓 Education & Training Development College 🚀 Innovation & International Collaboration College

Machine Learning Department Computer Vision Department Natural Language Processing Department High-Performance Computing Department Quantum Computing Department Generative AI & Foundation Models Center Reinforcement Learning & Agents Center AI Safety & Alignment Center AI for Science Center Robotics & Embodied AI Center

🤖

🔑

⭐ Richard Bellman

Chair

Researchers (real GeoVerse Lab members) 15

🔑

1919–2010

David Blackwell

Researcher

💡 Proved foundational optimality and convergence results for Markov decision processes (Blackwell optimality), establishing the rigorous mathematical conditions under which dynamic-programming-based sequential decision policies are provably optimal — extending and rigorizing Bellman's framework

🔑

1849–1936

Ivan Pavlov

Researcher

💡 Discovered classical conditioning — that an organism learns to associate a predictive stimulus with a delayed reward through repeated temporal pairing — the foundational associative-learning phenomenon that temporal-difference learning algorithms in RL directly formalize mathematically

🔑

1906–1991

Magnus Hestenes

Researcher

💡 Founded modern numerical methods for the calculus of variations and constrained optimization (conjugate gradient method, augmented Lagrangian methods), providing the gradient-based optimization machinery that policy gradient reinforcement learning algorithms directly apply to optimize policy parameters

🧑‍🔬

🔑

1941–1997

Harry Klopf

Researcher

💡 Proposed the 'hedonistic neuron' hypothesis — that individual neurons act to maximize local reward-like signals — directly inspiring Sutton and Barto's actor-critic architecture, which separates a policy ('actor') from a value-estimating critic

🧑‍🔬

🔑

1918–2016

Jay Wright Forrester

Researcher

💡 Founded System Dynamics, the discipline of building explicit feedback-loop simulation models of complex systems to predict and reason about their behavior — the direct conceptual and terminological ancestor of the 'world models' that model-based reinforcement learning agents learn and plan within

🔑

1923–2016

Lloyd Shapley

Researcher

💡 Founded the theory of stochastic games, generalizing Markov decision processes to multiple interacting agents with competing or cooperative objectives — the exact mathematical framework underlying multi-agent reinforcement learning

🧑‍🔬

🔑

1901–1990

Arthur Samuel

Researcher

💡 Built the first self-improving game-playing program (checkers) that learned by playing against itself and evaluating board positions via minimax search, coining the term 'machine learning' and founding the self-play tree-search paradigm that MCTS and AlphaZero-class systems directly descend from

🔑

1900–1970

Arturo Rosenblueth

Researcher

💡 Co-founded cybernetics with Wiener and Bigelow, formalizing the theory of goal-directed, feedback-driven behavior in machines — the founding theoretical framework for agents that pursue goals via perception-action-feedback loops, directly relevant to LLM-based autonomous agents using tools and feedback

🔑

1908–1984

Gustav Elfving

Researcher

💡 Founded the theory of optimal experimental design, determining how to choose experimental conditions (the 'Elfving set') to maximize the statistical information gained per experiment — foundational to Bayesian optimal experimental design for autonomous scientific exploration

🧑‍🔬

🔑

1927–2021

Peter Whittle

Researcher

💡 Developed the theory of restless bandits and Whittle indices, generalizing multi-armed bandit theory to more realistic settings where unobserved arms continue to evolve — foundational to modern exploration-exploitation algorithms balancing information gathering against exploitation of known-good options

🧑‍🔬

🔑

1927–1992

Allen Newell

Researcher

💡 Co-developed the General Problem Solver and means-ends analysis, the foundational method of decomposing a complex goal into a hierarchy of subgoals — directly foundational to hierarchical reinforcement learning's options and subgoal frameworks

🔑

1925–2021

Albert Bandura

Researcher

💡 Founded Social Learning Theory, demonstrating that organisms learn complex behaviors by observing and imitating others (the famous Bobo doll experiments) without direct trial-and-error reward — the founding psychological phenomenon that imitation learning and behavior cloning in RL directly formalize

🧑‍🔬

🔑

1924–1976

Daniel Berlyne

Researcher

💡 Founded the experimental psychology of curiosity, showing that novelty, uncertainty, and complexity themselves drive exploratory behavior independent of external reward — the foundational psychological theory that intrinsic-motivation and curiosity-driven exploration bonuses in RL directly formalize

🧑‍🔬

🔑

1925–2014

Harold W. Kuhn

Researcher

💡 Developed the theory of extensive-form games (game trees with sequential moves and information sets), providing the exact formal representation of sequential, multi-step decision-making under uncertainty that both game-tree search and sequential multi-agent RL directly operate on

🧑‍🔬

🔑

1919–1997

Yakov Tsypkin

Researcher

💡 Founded the theory of adaptive and learning control systems, showing how a controller can adjust its own parameters online based on observed system behavior — directly foundational to adaptive control-theoretic reinforcement learning agents that tune their policies in response to a changing environment