「active-inference」の検索結果

Bridge active inference theory with robot control using K-Scale's JAX/MuJoCo stack. Use when connecting predictive coding to locomotion policies, mapping KL divergence minimization to RL training, applying mean field approximation to robotics state estimation, or implementing sim2real as inference about future observations.

日本語の概要は準備中です。原文の説明を表示しています。

Principled action selection for autonomous agents using Expected Free Energy (EFE) under the Active Inference framework. Each candidate action is scored as G(n) = w_prag * (-E[p_n] * damping) + w_epist * (-Var[p_n] * novelty(n)), where agent beliefs are maintained as per-node Beta distributions and updated via Bayesian observation. Selection is a softmax over -gamma * G, making exploration an automatic consequence of uncertainty rather than a hardcoded epsilon. Precision weights w_prag and w_epist self-adapt via surprise history, shifting the agent between exploitation and exploration regimes without any manually tuned schedule.

日本語の概要は準備中です。原文の説明を表示しています。

An innate exploratory drive that guarantees no graph node remains permanently invisible to an agent population. After a warm-up period (step > 3), each agent fires a coverage scan with probability |unseen_nodes| / |all_nodes|. When triggered, the agent teleports directly to the globally unseen node with the highest belief uncertainty, works it, and returns control to the normal action-selection loop. Probability is self-annealing: it starts near 1.0 when almost everything is unseen and decays toward 0.0 as coverage saturates, requiring no manual schedule. The mechanism is "innate" (hard-coded into the agent step loop) as distinct from the "adaptive" EFE path, which requires local neighbors and learned precision weights. Together they form a two-layer exploration architecture: adaptive for locally-guided foraging, innate for topological completeness.

日本語の概要は準備中です。原文の説明を表示しています。