Towards (Open Source, Active Inference) Bioregional Biofirm Modeling

All right. Hello, everyone. It's November 3rd, 2024. I'm going to begin active inference stream number 6.1 as per usual with a GitHub push, and then let's jump right into the repo, and this should be an awesome stream. So thanks to those who are joining live. For those who are interested in bioregional modeling, open source active inference toolkits, and more, this should be quite exciting. So I'm excited to walk through where things have gotten over the last few days and talk about how we'll continue over the next 10 days plus plus plus. All of this is happening in the open source repo biofirm. I'm sure we'll return to this in a lot of different ways in the coming stream, and everything that I'm about to go through is the current state on biofirm. We just push the update. Let's reload to confirm. Awesome. Welcome, John. All right. There's two aspects of the biofirm concept that the current state of the repo and this stream are going to demonstrate. The first is the active inference component, which is an active inference-based multi-agent control framework. One agent deployed per ecological modality with differing levels of noise and controllability within a general homeostatic framework. All of this, of course, work in progress, simple where possible. However, I ran into a lot of fun ways to work with PyMDP over the last few days years. Using cursor 0.42. The second part of the biofirm research is going to reuse some code from Gen24. This is going to use perplexity LLM with internet access calls with an API to do some bioregional research and structuring and reporting for California counties. Okay, where are we going? And then let's jump first into the active inference component, stream. Okay. The goals of the stream are to first demonstrate a proof of concept for bioregional modeling using active inference frameworks, building on what Andrew Pachea and others have done. Second, to prepare, work towards preparing some open source bioregional templating for the coming applied active inference symposium. And then third, explore PyMDP methods and simulations and get a better sense for what's really happening with that ABCD after all. Okay. Detail agenda. Just to read the highlights, then we'll jump into one. One, bioregional modeling outputs and impact. Two, more foundational background. Three, we'll go into detail on PyMDP, partially observable Markov decision process, POMDP. Four, talk a little bit more about bioregional modeling, and then there's a bunch of discussion topics. All right. Here's where we're going with this simulation. So I'll walk through the visualizations first. Here is 10 ecological variables. This uses in the utils file. I'm going to jump around a little bit. So ask any questions in the chat and also look at it synchronously on your own time. ecosystemconfig.json is going to describe the variables as well as their relationships amongst them. I kept that part empty, but it does work. Each variable has an initial value, which is set at 50, and a constraint homeostatic range. Here for all the variables, it's set at 45 to 55. Then each variable has several important factors that determine its controllability. First is a trend coefficient. This does work. However, I've left them as zero for now. And then two main factors that we're going to be exploring a lot, which is the control strength and the noise standard deviation. So what's going to happen in ecosystem simulation is it is going to run a random setting and watch the ecosystem unroll with no control. And then it's also going to run an active inference simulation where depending on the controllability, the active inference agent is able to reenter those scaled controls back into the ecosystem. So to get started, first you run from the scripts folder for now, Python 3, and then startup.py. Startup.py is going to install into a Python virtual environment, the PyMDP package. And then you need to run source, VMV bin activate in order to get it going. You'll…