| Assignment Type | Lab — Case Study Analysis |
|---|---|
| Topic | AlphaGo Zero reinforcement learning through self-play |
| Date Submitted | 15 February 2026 |
AlphaGo Zero was configured to learn patterns and strategies without any prior knowledge of the game it was designed to play. What makes this approach particularly challenging is the human-like complexity of decision-making required for each move, as well as the limited availability of high-quality training data. Collecting datasets that effectively replicate expert human decision-making is often expensive, unreliable, or simply impossible. Even when such datasets are available, they can impose a performance ceiling on trained systems (Silver et al., 2017).
The first step in training was the removal of all previous knowledge, allowing the system to learn entirely through self-play. The program plays games against itself, and at each position a Monte Carlo Tree Search (MCTS) is executed using the most recent version of the neural network. Moves are selected according to the search probabilities generated by the MCTS. Once a game is completed, the rules of the game determine the winner.
The neural network takes the raw board position as input, processes it through multiple convolutional layers, and outputs two components: a policy vector representing the probability distribution over possible moves, and a scalar value representing the probability that the current player will win. The network’s parameters are updated to maximize similarity between the policy vector and the MCTS search probabilities, while minimizing error between the predicted outcome and the actual game result.
| Milestone | Result |
|---|---|
| After 3 days of self-play | Defeated previous AlphaGo version 100–0 |
| Previous AlphaGo achievement | Defeated 18-time world champion Lee Sedol |
| Training approach | Zero human examples — rules only |
| Conclusion | Superhuman performance without human guidance is fully feasible |
Given appropriate parameters and rule-based frameworks, the self-play approach can be applied across many domains:
This case study was one of the most thought-provoking assignments of the semester. AlphaGo Zero demonstrated that removing human bias from training data can actually lead to better results. The system not only surpassed human expert performance but also discovered novel strategies that human players had never considered in thousands of years of playing the game of Go.
The broader lesson is that reinforcement learning through self-play is a powerful paradigm for any domain where the rules are well-defined but optimal strategies are unknown. It also made me reflect on what “learning” really means — AlphaGo Zero learned more effectively than systems trained on human knowledge, simply by being allowed to explore without human constraints.
Silver, David, et al. “Mastering the Game of Go without Human Knowledge.” Nature, vol. 550, 2017, pp. 354–359.
Hassabis, Demis. “AlphaGo Zero: Starting from Scratch.” DeepMind Blog, 2025.