# CyphAI ## Short update on data sets, simulators, benchmarks, and tools --- # Litterature, Biblio, etc ---- ## Hybrid Systems and AI/Learning - Minor topic in mainstream AI/ML - Dominated by - Vision - NLP - Sentiment analysis - ... ---- ## HS-like topics in mainstream AI/ML - Time series analysis for anomaly detection, predictive maintenance - Reinforcement learning - Not much explicit "hybrid" consideration - Timing/causalities aspect limited - RNN, LSTM, ... - throw some memory loop in the mix, see if it sticks ---- ## Hybrid Systems Learning Tools - Some litterature - Couldn't find yet mature tools - HIT from early 2000s still functional though - Clever mixes of clusterings for mode inference and system identification within modes --- # Business as usual ---- ## Breach update - Version 1.10 with (slightly) better documentation - Interface with generic simulator/data - Work-in-progress: - coverage measures, set filtering for learning data sets analysis and preparation ---- ## Falsification Benchmark - Signal Temporal Logics (STL) complex requirements inspired from Volvo testing specifications - Multi-requirement testing (HSCC'22 paper to appear) ---- ## Signal Temporal Logic Monitoring - "New" library github.com/decyphir/RobonTL - Pure C++, stand alone - Online/offline monitoring - Python bindings in the works Note: several other existing STL-based tools exist in Python (e.g., RTAMT) --- # Work in Progress ---- ## Wordgen - Timed automata traces uniform sampler - Interface with Breach for constrained inputs - Proof-of-concept closed loop execution with environment - Sumo case studies ---- ## Sumo - Efficient traffic simulator - small scale toy problems (single intersection, grids) - large scale, realistic scenario (import open street map of Paris) - Preliminary experiments with (wordgen, Breach) ---- ## Diabete Type I - deceptively simple problem - simple formulation - but (human) biology in the loop - Started working with data from public.jaeb.org/datasets/diabetes - Looking for models, richer data - Local startup company (Diabeloop), trying to get in touch ---- ## Other Data Sets, Case Studies - paperswithcode.com - Time series worth digging in - see code tab in arXiv too - RL resources - OpenAI gym - benchmarks repos with trained agents (RL-zoo) --- # What next ? ---- # Tutorials - Breach (overdue) - STL monitoring/mining - TA (wordgen, BIP) --- ## Sharing code and experiments Shared group gitlab.com/cyphai Can add subgroup for - datasets - benchmarks - tools and/or code example - Paper artefacts - CyphAI ultimate tool suite ----