One-line definitions for terms that recur across this repository.
| Term | One-line definition |
|---|---|
| Active Inference (AIF) | A first-principles theory of perception, learning, and action: minimise (expected) variational free energy. |
Free energy F |
Tractable upper bound on surprise; F = E_q[log q(s) − log p(o, s)]. |
Expected free energy G(π) |
Future-oriented free energy under policy π; decomposes into epistemic (info-gain) + pragmatic (preference). |
| Generative model | The agent's hypothesised joint distribution p(o, s, π) over observations, hidden states, and policies. |
Hidden state s |
The unobserved cause of observations, inferred via Bayes. |
Observation o |
What the agent's sensors deliver at a timestep. |
Policy π |
A sequence of actions over policy_len future timesteps. |
Action u |
The single action sampled from the action posterior at this timestep. |
| Affordance | A label for an action (e.g. "UP", "DOWN", "STAY"); the set is called E. |
Likelihood A |
Matrix `P(o |
Transition B |
Tensor `P(s_t |
Preference C |
Vector of (log-)preferences over observations; the agent prefers higher-C observations. |
Initial prior D |
Prior over hidden states at t = 0. |
Posterior q(s) |
Belief over hidden states after seeing the observation. |
Softmax temperature gamma |
Inverse-temperature on softmax(-G) controlling action determinism (gamma=16.0 default in pymdp). |
| cadCAD | Discrete-time generative simulation framework for complex adaptive systems (cadcad.org). |
| radCAD | Modern reimplementation of cadCAD with the same surface; ActiveBlockference uses it under the hood. |
| Partial state update block | The cadCAD unit that owns one set of policies + one set of state-update functions. |
| Generative Research Team (GRT) | An LLM agent crew assembled to write reports; lives in GRTs/. Unrelated to AIF agents. |
pymdp |
Upstream Python package providing the canonical AIF primitives (pymdp.agent.Agent, pymdp.maths, …). |