This approach casts the choice of anomaly-detection algorithm as a sequential decision problem: a reinforcement-learning agent, featurizing windows of a time series with a Time Series Forest, learns a policy that dynamically selects which detector to apply, and it trains without ground-truth anomaly labels. Because no single detector is uniformly best across global, local, and clustered anomalies, dynamic selection matches or beats every fixed detector on F1 and outperforms a prompted general-purpose LLM. Students learn how model selection itself can be solved with RL, how a reward function stands in for missing labels, and why an ensemble-of-detectors-with-a-chooser can generalize across anomaly types.
This framework performs unsupervised anomaly detection on time series by dynamically selecting, rather than fixing, the detection algorithm. A reinforcement-learning agent, using a Time Series Forest…