# David Silver > 1976– · Computer Scientist, AI Researcher > > **Recorded contribution:** AlphaGo; AlphaZero; reinforcement learning; UCB algorithms ## How to use this dossier Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory. ## 1. Historical orientation David Silver studied reinforcement learning with Richard Sutton and became a principal research scientist at DeepMind and professor at University College London. He led the reinforcement-learning research behind AlphaGo and AlphaZero, systems that combined neural networks, search, and self-play to master games from experience rather than relying only on handcrafted evaluation rules. ## 2. The problem inherited Search in games such as Go has an enormous branching factor, while classic reinforcement learning struggled to estimate long-term value accurately enough in large state spaces. ## 3. The central contribution Silver led the AlphaGo and AlphaZero research program, integrating learned policy and value networks with Monte Carlo tree search and self-play improvement. ## 4. Reconstruct the mechanism 1. Use a policy network to prioritize promising actions instead of expanding every move equally. 2. Use a value estimate or rollout to evaluate positions without playing every branch to completion. 3. Run Monte Carlo tree search to balance exploration with exploitation. 4. Generate stronger training data through self-play and repeat the policy-evaluation cycle. ## 5. What changed downstream - AlphaGo defeated leading professional Go players and demonstrated a powerful learned-search combination. - AlphaZero showed that a shared self-play recipe could reach high performance in multiple board games with less domain-specific knowledge. ## 6. Attribution, limits, and uncertainty - The systems were built by large DeepMind teams and inherit core ideas from reinforcement learning, UCT, neural networks, and decades of computer-game research. - Board games provide exact rules and cheap simulation; the same approach cannot be assumed safe or data-efficient in an open, costly, partially observed world. ## 7. Reconstruction lab Build Monte Carlo tree search for tic-tac-toe, first with random rollouts and then with a learned value table. Compare node expansions and identify one state where the value estimate harms search. Record the policy prior, visit counts, value estimates, and final move at each turn so that improvement can be audited. Run self-play from a deliberately weak initial policy and explain how exploration supplies new evidence. Then introduce a biased rule implementation and observe self-play optimize the wrong game. The central first-principles lesson is that planning amplifies the model and environment it is given; it cannot repair an incorrect simulator merely by searching more deeply. ## 8. Evidence trail - [David Silver](https://www.ucl.ac.uk/engineering/computer-science/people/david-silver) — University College London - [Mastering the game of Go with deep neural networks and tree search](https://www.nature.com/articles/nature16961) — Nature - [Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm](https://arxiv.org/abs/1712.01815) — arXiv --- *Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*