A quick note: Ray Summit is coming up next week, as a free online conference
https://events.linuxfoundation.org/ray-summit/Ray is open source in Python (and now Java too) out of
UC Berkeley RISELab for more contemporary patterns of parallel, distributed processing at scale (Ion Stoicha, Michael Jordan, Dave Patterson, et al.). That work is especially focused on real-time applications, and agent-based AI. Think of it as a next-generation after Apache Spark, following more of the architecture for what Tensorflow has leveraged. There are simple annotations for Remote Functions (
task-parallel), Remote Objects (e.g.,
distributed key/value store), Remote Classes (
actor pattern), drop-in replacements for highly parallelized alternatives to `multiprocessing.Pool` and also parallelizing the `JobId` backend in
scikit-learn projects, plus performance improvements in
pandas,
TensorFlow, and
PyTorch. All of the above gets used as foundations for higher level abstractions such as `RLlib` (reinforcement learning), `Tune` (automl), and `Serve` (high-performance RL model serving integrated with Flask).
FWIW, I will be teaching the reinforcement learning tutorial on Tuesday 29 September. For this I've been developing a few new open source Python resources:
- Guide plus code templates on GitHub for how to build OpenAI `gym` environments that are optimized to work with RLlib
- An example of using a contextual bandit to optimize an investment portfolio (using 92 years of US markets data from NYU)
- An example of implementing a recommender system using reinforcement learning (using the "Jester" dataset from UC Berkeley) which adapts to changes in personal preferences and changes in item content
See you there!
------------------------------
Paco Nathan
https://derwen.ai/paco------------------------------
#GlobalAIandDataScience#GlobalDataScience