Belief Programming in Partially Observable Probabilistic Environments
summary
The work introduces P-BLIMP, a probabilistic belief-programming language with true/observe/infer statements that maintain belief states as probability distributions over hidden variables. It builds a weakest pre-expectation calculus that symbolically computes expected postconditions and proves soundness: the calculus yields the expected value at termination. A case study models a disease with 10% prevalence and a 95% accurate test, deriving an optimal decision threshold (one candidate is 1/2). Future direction is unifying probabilistic and nondeterministic belief semantics.
AI-written summary · about
related talks
28mPreference Learning from Minimal Human Feedback for Interactive
21mProgramming-by-Demonstration for Long-Horizon Robot Tasks
25mHigh-Level Mission Specification for Multiple Robots
15mAbstraction Refinement-guided Program Synthesis for Robot Learning from Demonstrations
29mSynchronous Programming and Refinement Types in Robotics: From Verification to Implementation
53mGeneralization in Robotic Behavior Cloning: Pitfalls and Promises
10mMIND MELD: Personalized Meta-Learning for Robot-Centric Imitation Learning
26mPAARL: An Interactive System for Policy-Aware Planning in Autonomous Agents(…)