The Multi Agent Network (MAN) is a nonstationary online learning heterogeneous multi agent system.

Agents are functions with annotated input and output signatures and which accept lists for every input. Most practical ML functions will sample one of the list elements or compute the mean of each input argument list, but the list can also have length zero. The function can output None for no output in sparse networks of experts. Agents might be off the shelf pertained NN's, datasets, environments, or arbitrary functions. Agents communicate via a pub-sub architecture, and connectivity can be queried and changed at internal to the network at runtime. Communication endpoints are uniquely identified by an agent ID, local endpoint name, and type.

Integrate with ML pipelines. The node neural network also allows defining inside-out interfaces allowing easy supervised learning, reinforcement learning, or arbitrary ML pipeline training for end-to-end differentiable models.

Online agents perform unsupervised learning on each update step or N steps. Some dynamically change their connectivity to other agents by inspecting the subscribable endpoints in the MAN.

Common reward: reward is one type of data. Most MAN's include a reward hub or subnetwork to allow sharing this signal semi-cooperatively. .

The executor also passes itself into the call function for agents that are common agents. The common agent can then look at All of the agents in the executors agent pool to see if it wants to form parent-child connections. The agent makes parent child connections by calling the dot at parent function on its child and calling the dot add child function on itself

You have the search for your own parents and children

Parent:Top hidden side Top

Child:bottomgrad:selfname bottomgrad:parent name