Select actions by minimizing Expected Free Energy in partially observable environments.
npx skills add https://github.com/thistleknot/skills --skill active-inference
Please help me install this Agent Skill. Skill: active-inference Source: https://github.com/thistleknot/skills/tree/main/active-inference Command: npx skills add https://github.com/thistleknot/skills --skill active-inference
Active Inference helps an agent decide what to do next when the true state is hidden and there is no reliable scalar reward signal to guide exploration.
Use this skill to build a small active-inference agent that models hidden causes of a bug and chooses the next tool action (like searching a file or running tests) that most reduces diagnostic uncertainty until you reach a confident conclusion.
High-intent search queries and answers about installing and using this skill.
Action selection under uncertainty without a scalar reward is achieved by active inference, which minimizes Expected Free Energy to balance information gain and preference satisfaction.
Active inference is a Bayesian decision-making approach for POMDP-like environments that uses variational state inference and Expected Free Energy to select actions that reduce diagnostic uncertainty.
Setting up a generative model for agent planning requires defining likelihood, transition, preference, and prior matrices to update beliefs from observations and drive uncertainty-minimizing actions.
Active inference applies to debugging workflows by driving probing actions such as targeted file searches or test runs based on which action most reduces diagnostic uncertainty before committing fixes.
Policy selection in active inference agents supports planning modes including flat inference, sophisticated inference, and MCTS, selecting policies by evaluating Expected Free Energy across possible actions.
Expected Free Energy minimization requires a valid generative model with defined matrices, making it unsuitable for environments where reliable likelihood, transition, preference, or prior distributions cannot be constructed.
shushuznSpecify Bayesian models in PyMC and fit them with MCMC or variational inference.
StepowskiEricUpdate belief states across competing hypotheses as new evidence arrives.
Lord1EgyptRun variational inference on annotated single-cell omics data matrices with scvi-tools.
plurigridCompute compression-progress intrinsic rewards with PyTorch or JAX.
