Definition
Probability matching is a behavioral pattern often observed in reinforcement learning and psychology, contrasting with optimal ‘maximizing’ strategies. Instead of always choosing the action with the highest expected reward, a probability-matching agent distributes its choices according to the underlying probability distribution of rewards. While suboptimal in stationary environments compared to pure exploitation, it can be advantageous in non-stationary settings where exploring different options helps track changing environmental dynamics. It serves as a baseline for understanding exploration-exploitation trade-offs.
Summary
A decision-making strategy where an agent selects actions with frequencies proportional to their estimated probabilities.
Key Concepts
- Exploration vs Exploitation
- Reinforcement Learning
- Stochastic Environments
- Decision Theory
Use Cases
- Modeling human behavior in psychological experiments
- Robust exploration in non-stationary multi-armed bandit problems
- Analyzing suboptimal learning algorithms