In situ model-based learning in PAMELA
summary
PAMELA is a Clojure DARPA-funded toolset that embeds machine learning into model-based programs: it automatically captures observations into MongoDB+RabbitMQ, selects algorithms (example: Baum‑Welch HMM) from model structure, trains on streamed data, and writes learned parameters back into the running program. Demo uses a Raspberry Pi plant with JSON plans, an HTN planner and dispatcher, and shows a coin‑toss HMM learning transitions from 1,000 flips (biased 0.9/0.1 and 0.8/0.2 cases work; performance falls below ~0.8). Temporal bounds and millisecond‑scale transition waveforms (5–7 ms) are learned as learning variables.
AI-written summary · about
chapters (16)
- 0:00Intro
- 1:12A brief History and Acknowledgement
- 2:56What is Pamela
- 5:27A model Based Program
- 7:44Pamela Learning Architecture
- 9:36Flipping a coin is a simple instantiation of a random value
- 10:49A traditional example machine learning problem
- 11:02Biased Coin Problem Automata Diagram
- 14:16Modeling the Biased Coin in Pamela-4
- 18:24Baum-Welch - Viterbi, eBH = 0.9 BT = 0.1
- 24:08Transitions in cyber physical systems
- 24:34Rolling Down a Hill
- 27:05Bouncy Switch
- 29:19The Plant Interface
- 31:13Plant Interface in Clojure
- 31:28Switch and Bulb Example
related talks
10mML-based Log Analysis for Faster Debugging
30mReconceptualizing Machine Learning for the Real-time World
43mToward Human-Centered Machine Learning
26mThe Feynman Machine: AI architecture in Clojure/Neanderthal
27mBuilding Mini Minds Our Adventures in Teaching Small LLMs to Reason
57mGood Practices for Applied Machine Learning
11mEffective Sequential Monte Carlo for Language Model Probabilistic Programs
46mArchitecting Efficient Agents with Semantic Routing and Mixture-of-Tools