IEEE CoG 2019 - Industry Day - David Silver (DeepMind)
summary
Combines reinforcement learning and deep neural nets to solve games end-to-end: DQN learns Atari from pixels and reaches roughly human level across 57 games; a Go system using policy and value networks plus Monte Carlo tree search defeats Lee Sedol 4–1; AlphaZero learns from self-play (8 hours to beat the original, 40 days to 100–0) and later beats Stockfish in a 1,000‑game match (155 wins to 6); AlphaStar uses attention/LSTM and a league of agents to beat pro StarCraft players 5–0 under chosen conditions.
AI-written summary · about
chapters (7)
- 0:00Introduction to RL and AI
- 1:39The power of reinforcement
- 5:05Atari and game benchmarks
- 11:26Alphago and the game of Go
- 24:59AlphaZero and general play
- 36:15Starcraft and AlphaStar
- 55:43Real world applications
related talks
40mWhat did AlphaGo do to beat the strongest human Go player?
49mML: Teaching machine to learn from its own mistakes (...)
56mCan I do AI?
26mAnkur Narang; Senior Vice President – Technology & Decision Sciences at Yatra at Cypher2017- Part1/3
44mReinforcement learning: A gentle introduction
21mWhen the machines out-smarts us
28mDeep Reinforcement Learning: theory, intuition, code
42mDawn of the planet of artificial intelligence