A curated collection of courses, lectures, and readable codebases for learning deep reinforcement learning. The curator says every resource was personally reviewed.
A two-stage adaptive training framework for GUI agents reduces interaction and improves performance by learning where to train and how far to roll out, using intermediate checkpoints and pivot-based one-step RL, with recoverable states to avoid redoing prior steps.
Reinforcement learning trains an agent by interaction to find optimal actions, aligning with Krafton’s game-making focus; it can train on game rules to compete with players or aid development, and it improves AI's ability to learn and decide in complex tasks.
Imagination-based multi-agent world models improve policy updates; CA-MAMBA shows substantial interaction efficiency and reward gains over baselines under equal budgets, with cost parity reducing overall learner cost, though some limits in the learned multiplier indicate cost estimates may not be the bottleneck.
A ML-focused tutorial guide detailing post-training methods, evaluation, and scalable RL, with source-verified content and practical simulations; includes reading lists per chapter.
The paper presents a reinforcement-learning method that trains an agent to act, reflect on trial experience, and write task-specific playbooks for future rollouts. It reports higher held-out benchmark performance than baseline methods.
The tutorial explains reinforcement learning through a robot navigating to a goal, covering rewards, discounted future returns, exploration, and Q-learning. It shows how Q-learning updates action values and why an action with an immediate penalty can still lead to a better outcome.
A project documenting a 2D standup robot (standai) trained via PPO. A 512x512 network learns to stand from supine/prone starts, using a rule-based teacher for distillation and end-to-end fine-tuning. Results cover multiple resolutions, seeds, and ablations, with a reproducible workflow in a GitHub repository.
The document argues for turn-level analysis over token-level credit when evaluating long-horizon LLM agents, notes variance from mixed-length turns, suggests a runtime critic as an anomaly signal to trigger retries, and questions calibration as policies drift and the cost of maintaining deployment-grade reliability.
The document describes a variant of the DQN algorithm where the online network selects the action and the target network scores it, mitigating over-optimistic high Q-values; mentions van Hasselt et al. (2016) and notes a discrete off-policy MLP implementation.