Please see here for a more thorough introduction to bagging and boosting algorithms. This is useful because it avoids overfitting to samples that can be easily classifiedĪnd instead tries to come up with models that are able to classify hard examples, too. When building the next tree, those samples that haveīeen misclassified before have a higher chance of being used to generate the tree. By combining the output of several smallĭecision trees, an ensemble learner (right) might end up with a higher accuracyīoosting algorithms start with a single small decision tree and evaluate how well PolicyMap (_map.PolicyMap)ĭeep Learning Framework (tf vs torch) Utilitiesĭistributed PyTorch Lightning Training on RayĪ single decision tree (left) might be able to get to an accuracy of 70%įor a binary classification task. RLlib Sample Collection and Trajectory Viewsīase Policy class (.Policy) RLlib Models, Preprocessors, and Action Distributions RLlib: Industry-Grade Reinforcement Learning Model selection and serving with Ray Tune and Ray ServeĮxternal library integrations (tune.integration) Workflows: Fast, Durable Application Flows
Pattern: Using ray.wait to limit the number of in-flight tasksĪntipattern: Unnecessary call of ray.get in a taskĪntipattern: Accessing Global Variable in Tasks/ActorsĪntipattern: Closure capture of large / unserializable objectĪdvanced pattern: Overlapping computation and communicationĪdvanced pattern: Fault Tolerance with Actor CheckpointingĪdvanced pattern: Concurrent operations with async actorĪdvanced antipattern: Redefining task or actor in loopĪdvanced antipattern: Processing results in submission order using ray.getĪdvanced antipattern: Fetching too many results at once with ray.getĭatasets: Distributed Data Loading and Compute Limiting Concurrency Per-Method with Concurrency Groupsīest Practices: Ray with Jupyter Notebook / JupyterLabĪsynchronous Advantage Actor Critic (A3C)