Affordance-based control (Fajen 2005, 2007)

The most commonly studied tasks in the ecological approach involve the perceptual control of actions such as interception and steering. These models all involve perceiving some variable and moving so as to null the discrepancy between a current value and an ideal value. However, none of these approaches involve the perception of affordances; specifically, none of them address how people work to keep the required corrections possible, given their action capabilities. Fajen (2005, 2007) proposes affordance-based control, an ecological research framework that brings these questions to the fore and leads to the discovery of new, affordance based control strategies that account well for the data and solve the problems of simple information-based control models. 

My current sense is that Fajen is absolutely correct in his assessment of the problems and has done sterling work developing an ecological solution. What follows is a brief description of the problems and his solutions; in the future I will blog some thoughts as I work to align my throwing affordance work with this framework.

Information-based control
The most common way to model the perceptual control of action is as a control law working to make a perceptual (typically visual) variable maintain a set value (see pretty much everything Bill Warren has ever done, for example). The most common example is the outfielder problem: you intercept the ball by moving so as to either cancel the optical acceleration (OAC) or to preserve a linear optical trajectory (LOT) of the ball. For OAC, the variable is optical acceleration; the control law is 'move so as to null the acceleration'. This basic approach has been applied to study the perceptual control of steering, interception of a target on foot or by hand, and braking to avoid a collision.

Brett Fajen (2005, 2007) identifies two key features of these theories. First, they all involve nulling an error, i.e. moving to as to make the current value of a perceptual variable the same as some ideal or goal value. Second, they all work with single variables; OAC only entails optical acceleration, for example. These are a package; the control law is working to keep a specifying variable at some target value. 

Fajen identifies two problems that come out of this. One is the growing literature detailing the use of non-specifying variables; the standard accounts have no room for such variables, but people do indeed seem to use them sometimes. They are still ecological variables, in the sense that they are structures in energy arrays created by the interaction of those arrays with the task dynamics; they simply don't specify the critical part of the task dynamic and instead just correlate to it. Ecological psychology is still struggling with how best to talk about the use of such variables. The second, and more important issue, is the problem of action boundaries. We need to control our behaviour so that the ideal/required state of a variable remains an achievable state, but not all errors can be nulled. For example, you could be running flat out and still not be able to cancel the optical acceleration if the ball is not catchable by you. The behaviour of optical acceleration is specifying that your running behaviour is not sufficient to catch the ball, but it did not specify that your running could never have been sufficient (or, conversely, if your running would indeed do the trick). Another way to frame the problem is the specifying variables controlled in typical models do not specify affordances.

Fajen proposes an alternative framework, which he refers to as affordance-based control. The basic idea is simple: instead of moving so as to null some error between an ideal and current state, he proposes people move so that the ideal state does not cross the action boundary separating possible from impossible requirements. To remain an ecological account, this implies that the ideal state, the current state and your maximum ability to create a state must all be perceptually available. 

Deceleration to a Safe Stop
Imagine you are driving a car, and you are approaching an intersection with a car stopped at a red light. You must stop your car just behind the other car without hitting it or without slowing to a crawl and taking forever to come to a stop. 

At each point in your trajectory as you slow down, there is an ideal deceleration that will bring you to a soft stop at the right place. There is also a current deceleration that is either greater than, less than or equal to the ideal deceleration. You can then also perceive both these things; the task dynamics of braking define both these things and both parts of the dynamic create information. 

If you do the standard error-nulling strategy, you will be constantly working to keep your perceived current deceleration equal to the perceived ideal deceleration. But people do not do this, because this might mean it takes too long to cross the space. Drivers typically wait for a bit until beginning to decelerate, and then spend time either under- or over-shooting brake requirements. The question then becomes, how long can you wait and when do you make a correction? The answer depends on road conditions, the strength of your brakes, etc. In all cases, without acting, the ideal/required deceleration will eventually outstrip the maximum possible deceleration and you will crash. What people therefore need to do is to work to ensure that the ideal deceleration does not exceed the maximum possible deceleration, so that the task continues to afford safe stopping. No standard account of visually guided braking allows this: the variables talked about specify whether or not your current braking is sufficient, and not whether the required braking is possible. 

The affordance-based control strategy proposes that you must calibrate the perception of ideal deceleration so that it is perceived in units of maximum possible deceleration. Instead of perceiving that ideal deceleration is 25m/s/s, we need to perceive that it is, for example, 50% of our stopping power. We then move to keep it below 100%, and exactly how we do this is up to us and the local constraints - you can be conservative and brake often, or riskier and brake later ("Adjustments are not made around an ideal state but rather within a safe region"; Fajen, 2005, pg 727). 

Here's how you do this, using only information. First, you perceive your current deceleration (e.g. location of the brake). Second, you perceive the ideal deceleration (specified by Global Optic Flow Rate / 2 * tau). If you are correctly calibrated (i.e. perceiving ideal deceleration as a percentage of your actual maximum deceleration) then the values of these two variables will always drift away from one another. As braking is increased and you decelerate faster, perceived ideal deceleration will decrease; as you lift off the brake and deceleration decreases, perceived ideal deceleration will increase again. Critically, this is not true if you are incorrectly calibrated. If you are perceiving ideal deceleration in the wrong units (e.g. you have just hooked a large trailer to your car and are still acting as if the brake will work the same as without the trailer) then perceived deceleration and perceived ideal deceleration will drift towards each other. So in this case, the perception of two perceptual variables and the relationship between them is enough to produce calibration and therefore successful affordance-based control; if you try to move so as to keep ideal deceleration below maximum possible deceleration, you will succeed if it is, in fact, possible.

Implications
So much good stuff comes out of this account:

1. Perceptually guided actions become about affordances again
2. Perceptual learning becomes a critical part of the analysis, and specific predictions can be worked out about the consequences of using specifying and non-specifying variables.
3. Individual variation in behaviour falls out of the analysis; there are many ways to keep an ideal state within a wide safety zone and not cross an action boundary
4. The whole thing remains informationally based, but emphasises the use of multiple variables and the relations between them as well as the need for calibration

Right  now I'm still getting my head around the details, and in particular I'm thinking about how to adapt my work on throwing affordances to align with Fajen's approach. One thing I'm missing in my throwing studies right now is a measure of maximum throwing speed for each individual, although I have some data showing that the effects of various manipulations do depend on whether people are throwing fast or slow. I'm going to start thinking about how to get this to all work, though.

Summary
I've known Brett for years, meeting every now and again at talks and conferences. I've always liked his stuff, but until now I've never quite realised what it was he was up to. Turns out I've been missing out (so thanks to a reviewer on my throwing paper for the offhand mention!). I'm looking forward to trying to apply this framework to my own work, to kick the tires of this account and see what it gets me. If you are interested in the visual control of action, I thoroughly recommend these papers.

References
Fajen, B. R. (2005). Perceiving possibilities for action: On the necessity of calibration and perceptual learning for the visual guidance of action. Perception, 34(6), 741-755.

Fajen, B. R. (2007). Affordance-based control of visually guided action. Ecological Psychology, 19(4), 383 - 410.