Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

all right September 9th 24 and Andrew go for it um hi everyone so uh as we just briefly I know some is coming in uh as we just briefly discussed uh since today is another session where we are not directly talking about any particular the chap chapter of the textbook but rather um just talking about applying active inference more generally um we agreed that I would spend just a little bit of time here uh giving sort of a quick rundown of how to construct an active inference agent in this case it'll be specifically uh a PDP agent meaning we're working with um well in this case we're working with discrete time and discret observations um as opposed to continuous time models which you'll see in chapter 8 of the textbook um and I kind of just want to focus on just from scratch like how do we think about constructing an agent um I won't be touching on how to construct an environment um but but that's the kind of thing that you know you would you would bear in mind whenever kind of coming up with an idea for a simulation or experiment you want to run um and actually on that note Daniel is it possible for me to um share my screen by chance yep go for it awesome thank you so this was the um and if anyone wants this link they can find me on GitHub a p hea just like my last name on the zoom uh and the repositories called ic2 S2 active inference tutorial um this is material that I presented back in July at University of Pennsylvania um I basically uh gave a tutorial on AC inference how it might be applied to social sciences specifically in uh constructing multi-agent simulations so um and is this uh is this visible to everyone the slides here yeah go for it for you know five to 10 minutes however much you just want to footnote and then let's see what people want to continue with definitely sounds good um so in this tutorial along the way I you know we're introduced active inference there's some nice um explanations of different concepts from the textbook Etc talking about uh the basy brain hypothesis General ideas how Bas rule kind of plays out in active inference and eventually we reach where uh we're going to use the pmdp package um which has been developed to make discrete time AC inference agents in Python and um there are a lot of updates being made to that package right now if you were to like use a git clone to download it directly there's a lot going on in there but for the time being I'm going to just be using the the PIP install um Pi Pi version uh it's a little bit more kind of uh packaged up comes as more like a single unit so um whenever we construct active inference agents uh the starting step should be in my view starting step should be what are the hidden states that the agent has in their generative model like what are they trying to infer uh second thing is what are the observations that the agents will encounter in the world and then three what are the actions uh or controls or policies those three terms are sometimes used interchangeably they don't necessarily mean exactly the same thing um but for now we'll just go through with all of those we'll call them actions uh what actions can the agent commit uh so so we can view that kind of like like this like the this is the agent that we be constructing um all of these things play out it's called an action perception Loop um let me see if I can just okay hopefully that's visible enough to everyone um so so that we'll we'll start at the bottom of this action perception mod um first observations there are two kinds of observations that this agent will receive at any given time step um quick note that this entire action perception Loop plays out at every time step we're working in discrete units of time and so you know whether it's from second to Second minute to minute millisecond a millisecond however um all of the units of time are the same uh here the agent receives one of two observations one of them is related to uh an improve Improvement if something improves or if it…