GNN for AgentMaker for PyMDP for Active Inference Biofirms for Bioregionalism... for Ants?!?!?

All right. Hey everyone. It's active in France stream number 007.1, and it should be a great and exciting stream. It's November 7th, 2024, 7-11, depending on where you are. And I'm going to begin the stream with a GitHub push that we are going to spend the next several hours probably unpacking. Let's commit it, push it, and begin what will be a very fun exploration into the Active Inference Institute's branch of PyMDP. While it's writing, let's jump into the show. Here's where we're heading. We're heading into Shapley value-analyzed homeostatic satisfaction, expected free energy, belief accuracy, and control efficiency analyses for multiple differently parameterized PyMDP agents, base, risk-averse, exploratory, and balanced on those measures. Being able to analyze combinatorics like the synergies of their belief accuracy, control efficiency, expected free energy, and homeostatic satisfaction. We can look at the report and explore different topics like, with respect to those variables, what's the mean Shapley contribution, coalition performance, best coalitions. This is all combinations of 0, 1, 2, 3, and 4. So for example, for expected free energy, we see that the 4-unit coalition is the best. Whereas for homeostatic satisficing, 0, agent 0 working alone does the best. Underneath these coalitions and their combinatorics are specific agents that are logged in individual experiments. And here's some of the plots and how that looks for each individual agent. This is a given active inference agent. Here we have policy entropy over time, action selection distribution, belief observation cycle, belief accuracies over time, homeostatic state distributions and trajectories, state transitions, policy selection probability, policy updating, decrease, maintain, and increase, belief dynamics through time, hidden state beliefs, action frequencies, policy convergences, belief convergences, correlations between beliefs and actions, and action confidences. So what that Shapley analysis is doing is running all the combinatorics under the hood. This is going to be a wild ride. So let's confirm that the GitHub has pushed. I'll put the code link in. Welcome to the 007 fork. In the live chat, I will look forward to everyone's comments. I am going to ensure that we are pushing the right way. It still has not loaded. Let's begin though with a 007 here. Taking a little bit of time to upload on GitHub. Okay, here we go. A little bit more context before we dive into what may be some absurdly technical, but really useful and interesting things that have happened over the last few days and hours. Here's the stream. GNN for agent maker for PyMDP for active inference bio firms for bioregionalism for ants. Let's decompress that title. Generalized notation notation. GNN. As semantic middleware for agent maker. A custom robust modular agent engineering framework. Using PyMDP. Python language package. For Markov decision processes with methods for active inference. And free energy based inference methods for bio firms. As proposed by John Klippinger as a paradigm for firms in the age of. Bioregionalism as a mode of life. Coextentialism for species like ants, people, and beyond. Here's some live stream code names and an only lightly redacted. Preludium. For your priors only. Predictions are not enough. The agent who minimized me. On her majesty's statistical service. Quantum of precision. Live in lead update. Professor Friston. Do you expect me to reach the optima? No. I expect you to optimize. Just like they wrote. At the end of the 2022 textbook. When the last words were. Ultimately we are confident that you will continue to pursue active inference in some form. Is that agents generative model. An MDP. Yes. An MDP. A PO. No. MDP. As if. The media were not already the message for you. This is a just-in-time live stream going through raw open source code. I know that there are bugs. As well as typos. Incorrect or ineffective calculations. Other errors. This is…