Textbook Group meeting for Parr, Pezzulo, Friston 2022 .

Hello, it's March 22nd, 2023. We're in Cohort 2 in our second discussion of Chapter 9. We're going to jump right in to anything that people are thinking about related to Chapter 9 or just in general. So whomever would like to bring something up, go for it. Yeah. So I have a question that you said last time, I don't know whether it was in what context it exactly was, that you would say, we talked about, or you talked about affordances, and you said that policies are not affordances. And then I was wondering what then affordances are, because I understood policies are not affordances because they are in the future. But then I was thinking what that should mean, because the present is a point in time, infinitely small, which would imply that affordances doesn't exist. Or how would you see this? So if one thinks that an affordance is, if we are in one state, and the policy is a sequence of states in the future, but then even if we have this embedded states, I can be in the state of learning, for instance, active inference, which also implies some actions in the future. So I was a little bit confused about this notion, and why you would think that policies are not affordances. Maybe I misunderstood. Thanks. Thanks. Ali, want to give a first thought, and then I'll be happy to add something. Well, yeah, actually, well, in fact, in active inference literature, there are some divergence between the way terms used in this context, as compared to their other connotations or other meanings in some other context, such as reinforcement learning and so on. So basically, in active inference, by policy, they generally mean a sequence of actions. But for instance, in aria learning policy is defined as just a single action. And affordance, again, is a kind of problematic term because we have different notions of affordance, for example, in ecological psychology or in some other areas. But to the best of my understanding, in active inference, affordance refers simply to a single action in some cases. But in some other cases, it can take on some other additional layers of meaning. So, for example, we can affordances or pragmatic affordances and they somehow conflated with each other and they can be confusing to delineate which notion of affordances is referred to here in this context. So, yeah, those are pretty much the basic or at least some of the points that can help delineate those two concepts, affordance and policy. But I'm sure Daniel would have some other thoughts on that as well. Awesome. Yes. A lot of good points. So, big picture, there are divergences between how the terms are used in ACTIMF and in reinforcement ecological psychology. We hope and basically believe that there's coherence in how the terms are used, but that doesn't guarantee coherence within and across areas which themselves are not necessarily internally or externally consistent. So, policies are sequences of actions over a given time horizon. Affordances are the instantaneous single action possibilities described by the E variable. So, map not territory. If you're in a maze and you can be going up, down, left, right, then those are your affordances. And over a time horizon of four, you could have policies of a time horizon of four with all the combinatorial possibilities of up, down, left, right, to the fourth power. Then there are those sometimes informally applied statements like a pragmatic affordance or an epistemic affordance, which is like when action is being taken that is oriented around epistemic value. But this is kind of like a modifier, like it's a risky affordance or it's a, you know, fun affordance, not necessarily like a special subtype in a formal way. And it's described by the E variable. You could have a nested generative model. So, like Livestream 42 with a slam, simultaneous localization and mapping. At the lower level, the robot has up, down, left, right affordances. At the higher level, the policy selection, the affordances have to do with…