Oak Lab

Prior work

• Talk

  The OaK Architecture

  A concise tour of the OaK agent design—how perception, model, policy, and value fit together for learning in big worlds.

• Paper

  The Alberta plan for AI research

  How the Alberta groups frame AI: continual learning, big worlds, and building agents that keep improving over a lifetime—open to all who share the direction.

• Paper

  The Big World Hypothesis and its Ramifications for AI

  The world outgrows any agent: prior training is never enough, continual learning matters, and cheap, reliable learning algorithms are essential.

• Paper

  SwiftTD: A Fast and Robust Algorithm for Temporal Difference Learning

  A temporal-difference method aimed at fast, stable credit assignment when learning from long, noisy streams of experience.

• Paper

  Step-size optimization for continual learning

  Meta-gradient methods like IDBD tune step sizes toward the true objective; common optimizers use heuristics that can drift away from better step sizes—relevant for lifelong learning in neural nets.

• Paper

  Reward-respecting subtasks for model-based reinforcement learning

  Subtasks and options defined to respect the main reward so discovered temporal abstractions help planning with a world model, not only auxiliary or bottleneck goals.

• Paper

  The quest for a common model of the intelligent decision maker

  A cross-discipline view of the intelligent agent: inputs, goals, and internal structure for perception, world models, and choice—naming what psychology, control, and AI already share.

• Paper

  Reward is enough

  A position paper: maximizing reward is enough, in principle, to yield perception, language, and other hallmarks of intelligence through reinforcement learning and trial-and-error experience.

• Paper

  Planning with expectation models

  When a value function is linear in features, planning with a learned expectation model can match planning with a full distribution model—plus sound model-based policy evaluation in stochastic settings.

• Paper

  On the role of tracking in stationary environments

  Tracking the best policy can beat any converging algorithm even on stationary problems (e.g. the Black and White problem, computer Go), with links to metalearning and IDBD for step-size adaptation.

• Paper

  Scalable Real-time Recurrent Learning using Columnar-constructive Networks

  Constraints on RTRL so online recurrent learning scales linearly in parameters—no extra gradient bias from truncated BPTT-style shortcuts, at the cost of some network flexibility.