RT @LazyDynamics: Backed Trojan Robotics (Team 24090) at the FIRST® Tech Challenge European Premier Event in Eindhoven (July 1–5)  . They h….

Some more refs: for implementation, (toolbox), and (AIF planning as inference), and (company). Also check out , and .

(6) Finally, FEP is more than a pretty good idea as it can be derived from first principles by information theory, see e.g., blog at plus refs. An AIF process avoids ad hoc design choices often found in man-made AI algorithms.

(5, cont'd) The advantages of RMP implementation are underestimated. Through RMP implementation, AIF agents keep working if you cut some energy supply, if you miss some observations, and if you need to interrupt to make a decision. See toolbox.

(5) An AIF process relies 100% on VFE minimization; hence, it can be fully implemented as reactive message passing (RMP) on a graph. This enables parallel, spatio-temporally distributed computation, making AIF agents robust to fluctuations in resources like time, data and energy.

(4) In an AIF process, the model spec is fully separated from the reasoning process, which is 100% based on variational free energy minimization (VFEM). An AIF process is therefore _explainable_ (by model assumptions) and _trustworthy_ (by Bayes-optimal reasoning process).

The more well-known advantages are (1) active data selection, (2) no need to specify a reward function, and (3) no need to extend the reward function with explorative incentives. Some of the lesser-known advantages:.

Agreed with @fchollet on FEP (, but FEP is more than a pretty good idea, and there are more benefits to realizing an agent as an active inference (AIF) process beyond active data selection. I will mention a few below:.

RT @ReactiveBayes: That is 🤯.

RT @ReactiveBayes: Probabilistic modeling in Julia just got smoother. With RxInfer + GraphPPL, you write models like regular code — loops,….

RT @mjdramstead: New blog post by @noumenal_labs: “Grounded rewards in the era of experience: A commentary on ‘Welcome to the era of experi….

Planning-as-Inference may become a reality for active inference agents! See our paper . (4/4).

. in a generative model that is augmented by both goal-directed and epistemic priors. We view this work as a new starting point for developing AIF agents that perform all processing, including planning, through straight VFE minimization. (3/4).

Unfortunately, computing EFE and ranking policies by EFE is computationally very challenging. In a new paper at we show that EFE-based planning can in principle be realized by regular Variational Free Energy (VFE) minimization . (2/4).

New paper alert! Active Inference (AIF) agents score policy candidates using a cost function called the Expected Free Energy (EFE), which has many desirable features, such as a parameter-free balance between information-seeking and goal-driven behavior. (1/4).

RT @LazyDynamics: Bayesian Inference in the browser? Yup. With new RxInfer TypeScript SDK, enabling real-time, client-side probabilistic re….

RT @ReactiveBayes: Did you know? 🤔 You can optimize parameters in RxInfer.jl using Optim.jl! Check out our latest example: .

RT @ReactiveBayes: New in RxInfer: VAEs for dynamics modeling 🚀. Elegant integration of AutoEncoderToolkit.jl VAEs with temporal prediction….

RT @ReactiveBayes: Excited to announce: @LazyDynamics is funding a new client-server infrastructure for RxInfer.jl to make probabilistic a….

RT @ReactiveBayes: Turns out .ml doesn’t just stand for machine learning - it also means maybe lost :/. But hey, every project needs its "w….