A **learning race** is a single training run that uses Counterfactual Regret Minimization (CFR) to iteratively converge toward a Nash equilibrium strategy for a given game configuration. Each race starts from a blank strategy profile and repeatedly traverses the game tree, accumulating regrets and updating the average strategy until the target iteration count is reached.

## How It Works

When you start a learning race, the system launches a CFR solver that alternates between two phases: **traversal**, where it walks through game decision nodes and updates regret values for every action, and **strategy update**, where cumulative regrets are translated into an improved strategy profile. Over thousands or millions of iterations, the strategy converges toward equilibrium, the point where no player can improve by unilaterally changing their play.

Training progress is tracked through periodic **checkpoints** that snapshot the current strategy and evaluate exploitability. These checkpoints let you monitor convergence in real time and compare the quality of different training configurations.

## When to Use Learning Races

Learning races are the core mechanism for producing equilibrium strategies. You create them from an equilibrium's settings page, configure CFR parameters (discount intervals, pruning thresholds, parallelism), and let the system iterate. Once a race reaches acceptable exploitability levels, the resulting strategy can be loaded into a solver and served to agents for real-time decision support.