02
Praxis Studio Self-Play Reinforcement Learning & MCTS
Reinforcement Learning · Monte Carlo Tree Search · AlphaZero Architecture
Teaching a machine to think through chess. An experimental research environment exploring how AI systems learn through self-play, neural evaluation and search.
Surreal chess pieces suspended over a warped board — search, strategy, and self-play
Teaching a machine to think through chess.
Praxis Studio is an experimental research environment exploring how an AI system can learn to play chess through self-play, neural evaluation and search.
The project takes inspiration from the AlphaGo Zero and AlphaZero research programmes, but treats the system as a research instrument rather than simply a chess engine. The goal is to make the learning process visible: what the model believes, where it searches, how those beliefs change, and how repeated self-play gradually produces stronger decisions.
Instead of encoding thousands of chess heuristics by hand, the system starts with the rules of the game and learns through repeated interaction with itself. A neural network produces two signals from a board position:
- Policy (P) — which moves appear promising.
- Value (v) — how favourable the current position appears.
Monte Carlo Tree Search then uses these predictions to decide where to spend computational effort. Promising moves receive more simulations, deeper branches are explored, and the resulting search statistics become training targets for the next iteration.
Live interactive self-play and evaluation engine · praxis.knolink.co.za
01 · The Learning Architecture
MCTS Multi-Path Expansion & Pruning
s₀ (Initial)
Ply 0
1.00
1600
+0.12
Initial root position evaluated by neural network f_θ(s₀). Generates baseline prior distribution across all legal candidate actions.
Clicking any candidate move above reveals its accumulated visits and neural valuation. PUCT continuously balances high-probability exploration with deep simulation down lines where expected win-rate Q(s,a) remains elevated.
The search is the experiment
One of the central ideas behind Praxis is that the model should not simply output a move.
It should reveal how it arrived there.
The interface exposes the search tree as it expands across the board, showing candidate moves, visit counts, policy probabilities, value estimates and the changing distribution of the search.
A move such as e4 is therefore not just a prediction. It becomes the visible result of thousands of competing possibilities being explored, evaluated and discarded.
PUCT selection rule
Score(s, a) = Q(s,a) + c_puct · P(s,a) · (√N(s) / (1 + N(s,a)))
Where Q(s,a) is accumulated action value, P(s,a) is the neural prior, N(s) is total state visits, and c_puct controls exploration.
The important idea is the tension between what the network already believes and what the search still needs to investigate. High-value moves attract search because they look good. High-prior but under-explored moves receive an exploration bonus. Over many simulations, the tree concentrates computation around the lines that appear most informative.
From search tree to learning signal
The search does more than select a move. It produces a richer target than the network's original prediction.
The policy head may initially assign probability across many legal moves. MCTS then transforms that uncertain distribution into a search-informed policy based on repeated exploration.
board position (s) → search policy (π) → game outcome (z)
The network is subsequently optimized to become better at predicting both:
- 1. The network guides the search.
- 2. The search generates experience.
- 3. Experience improves the network.
- 4. The improved network guides a better search.
This self-improvement loop is the foundational breakthrough demonstrated in AlphaGo Zero and AlphaZero.
What the studio makes visible
Praxis Studio is designed around the telemetry of learning. Rather than hiding training behind a single model checkpoint, the interface exposes the system's internal process:
02 · Internal Geometry
Praxis Latent Space Geometry
Experimental visualization of internal representation geometry, activation energy and positional embeddings.
Encodes control of d4/e4/d5/e5 squares. Activates strongly during early central tension structures.
Research Lineage
Praxis is situated within a lineage of systems that progressively moved away from handcrafted game knowledge:
Deep neural networks combined with Monte Carlo Tree Search, using policy and value networks to evaluate positions and guide search (Nature).
Removed human gameplay data entirely, learning from scratch through pure self-play reinforcement learning.
Generalized the architecture to chess, shogi and Go with a single shared algorithm and zero game-specific heuristics.
An experimental implementation and visualization laboratory making the internal dynamics of search, representation, and self-play telemetry observable.
The Research Question
“If we can watch an AI search through possible futures, can we better understand how learning changes the way it searches?”
Chess is the environment. Self-play is the data generator. The neural network is the evaluator. MCTS is the search mechanism. Telemetry is the microscope.