Matineh Shaker, Artificial Intelligence Scientist, Bonsai
summary
Concept Network Reinforcement Learning decomposes hard RL tasks into shallow trees of reusable sub-concepts (policies) and selector nodes, and shows industrial applicability on OpenAI Gym Lunar Lander and a robotics pick‑and‑place stacker. Subpolicies train with policy gradients / TRPO, the selector uses a DQN value network, and transfer reuse reduces reward sparsity. Experiments report roughly order-of-magnitude faster convergence versus a flat baseline, produce a working pick‑and‑place pipeline (suboptimal but successful), and validate reusable concept primitives.
AI-written summary · about
chapters (4)
- 0:00Start
- 4:31Markov Decision Processes
- 5:17Exploration versus Exploitation
- 9:48Actions that an Agent Can Make
related talks
01h 17mTraining intelligent game agents using deep reinforcement learning
20mProject Bonsai: How Microsoft is Using Ray and RLlib Inside a Machine Teaching Service
01h 35mDeep Robotic Learning
28mDeep Reinforcement Learning: theory, intuition, code
26mAnkur Narang; Senior Vice President – Technology & Decision Sciences at Yatra at Cypher2017- Part1/3
31mDesigning game AI with reinforcement learning
31mLayla El Asri, Research Scientist, Maluuba, NYC 2017
01h 11mFoundations of RL