In the real biological system, I would expect a set of fast and slow acting mechanisms to minimize free energy. Fast acting mechanisms are globally aware with low information through it. Slow acting mechanisms still have a slow information transfer rate but converge over their longer period of time. By fast acting, I mean the brain can manipulate the lights in the code for a recurrent energy processing unit.

The energy nodes themselves have an energy-based model of their probability distribution. Rather than set constant parameterized limits, The energy-based models are penalized for divergence from their early developmental state. A clear energy-based model serves as a loss function for connected nodes

Train level should be part of the step function for you by fast training self organizing maps. Slow training processes like gradient descent should wait for the specialized train function.

Since a energy-based model on the node energy organ allows to define a cost function, the cumulative cost rollout is minimized by gradient descent. Every time step Becomes it's own batch where multi step future rollout is then performed (with multiple timesteps) and energy minimization is burned into the weights

Batch

x > y y y y

x x > y y y y

x x x > y y y y

min sum(-log EBM(sum(y, axis=1)), axes=0)

Individual organs also minimize their own free energies. Muscular free energy is the difference between demanded and available energy. Ocular and cochlear free energy are also given by the difference in demanded and available energy. This time "demanded" Energy refers to that required for nervous signal transmission. In fact demanded and free energy is such a widely used term, then it should be available as public attributes of the base Oregon class and serve as the indicator of stress (pain with directly in code by rapidly and increasingly demanding energy. Which produces free energy, which is motivational he aversive, and which is minimized for by the energy management system by putting the energy system into a state of stress)

I would like to be able to incorporate osteopathic free energy perhaps by divergence of the joint state.

Some brains even when done developing will still have high levels of free energy. Those with low average brain and body free energy will be selected to carry offspring agents which will develop while sharing some energy with the host. I may also introduce staged morphological development here

To do

Make sure node's received the incoming edges and outgoing edges — even when the node dictionary doesn't yet have a complete collection of those organs

Penalize high amounts of unchanging energy components

Training consumes energy and threshold functions determine when to train. Every organ has only a fixed amount of training energy however.

Training takes place during the step function. Individual organs should be able to nullify the organisms simulation loop delay. The brain may have a special function to detirmine when it trains which asynchronously runs gradient descent while locking peripheral nervous activity but still continuing energy processing and training of energy nodes. Also, the brain can still administer hormone vector bases during sleep

Relatively deterministic reactions proceed throughout energy conversion nodes. They are only altered by training and existing vector components — some of which may represent hormones.

Separate energy basis receptor and emitter from root energy nodes.