A systematic evaluation of the evidence for perceptual control theory in tracking studies

Highlights

• Tracking entails control of perceptual variables to referent states.

• Perceptual control models capture parameters and simulate individual performance characteristics in tracking.

• Hierarchical perceptual control models can support complex motor behavior

• Future research in perceptual control should focus on model comparisons, motor learning, and application in embodied robotic control.

Abstract

Perceptual control theory (PCT) proposes that perceptual inputs are controlled to intentional ‘reference’ states by hierarchical negative feedback control, evidence for which comes from manual tracking experiments in humans. We reviewed these experiments to determine whether tracking is a process of perceptual control, and to assess the state-of-the-evidence for PCT. A systematic literature search was conducted of peer-review journal and book chapters in which tracking data were simulated with a PCT model (13 studies, 53 participants). We report a narrative review of these studies and a qualitative assessment of their methodological quality. We found evidence that individuals track to individual-specific endogenously-specified reference states and act against disturbances, and evidence that hierarchical PCT can simulate complex tracking. PCT’s learning algorithm, reorganization, was not modelled. Limitations exist in the range of tracking conditions under which the PCT model has been tested. Future PCT research should apply the PCT methodology to identify control variables in real-world tasks and develop hierarchical PCT architectures for goal-oriented robotics to test the plausibility of PCT model-based action control.

Introduction

Tracking underpins many human activities and skills, such as reaching, manual tool use, and driving. Manual tracking behavior, in the simplest case, is a process of dynamic visual or tactile error correction; for example, keeping a car in-lane while compensating for bends in the road and disturbances such as wind. This view sees the human operator as a servo-control mechanism (Craik, 1948) which can be characterized and modelled by a transfer function (Navas and Stark, 1968; Noble et al., 1955; Poulton, 1952a). A transfer function is an equation that allows calculation of the range of outputs from input values. In a typical servo controller the set-point, or reference (goal state), is assumed to be specified exogenously by the target (the curvature of road in the above example). An alternate view was proposed by William Powers that within organisms, reference points are set endogenously such that the system controls perceptual input to a goal state determined by the organism itself rather than by its environment (Powers, 1973). In the driving example, this might equate to the driver having the goal maintain road positioning. The endogenous reference enables control systems to simulate purposeful, goal-oriented behavior via negative feedback control.

Powers’ theory, perceptual control theory (Box 1) was introduced in two papers in 1960 (Powers et al., 1960a, 1960b), and later expanded in the book Behavior: the control of perception (Powers, 1973). In PCT, perceptual states are maintained at reference points via negative feedback processes (Powers, 1978; Powers et al., 1960a). The reference point specifies the desired state in which an individual aims to maintain a perceptual variable (e.g. the distance between the individual and the next car ahead). For each perceptual variable, a comparison of the reference point and the present actual state of the perceptual variable (input signal) yields an error term. Outputs are varied dynamically to reduce this error to bring the input signal into alignment with the reference state for that perceptual variable and keeping it there. Complex behavior emerges from a hierarchy of control units, each controlling a different perceptual variable. The hierarchy operates as a two-way cascade. Bottom-up projections between hierarchical levels carry increasingly integrated perceptual information to higher level control units. The outputs at each level set the reference values for units in subordinate levels via top-down projections. Output signals from the lowest hierarchical level are translated into movement via the effectors. These movements produce effects on environment, and consequently, alter low-level perceptual inputs via environmental feedback. Learning is proposed to be enacted by a reorganizing system which projects to units in the hierarchy (Powers et al., 1960a). When the error signal within a control unit persists above a certain threshold the reorganizing system will alter the structure of the hierarchy and the parameters of its control units in a trial-and-error fashion until the error reduces and control is re-established.

The PCT theoretical formulation conforms to contemporary neurophysiological evidence regarding the structure of the CNS as the perceptual hierarchy mirrors the known hierarchical organization of the peripheral nervous system and sensorimotor cortices (Iwamura, 1998); the bidirectional cascade acknowledges the reciprocal connectivity between functional areas that supporting recurrent activation (Sporns et al., 2002); and reorganization has clear parallels in neural plasticity, synaptogenesis and functional localization (Dayan and Cohen, 2011). Moreover, PCT provides an interdisciplinary framework (Carey et al., 2014) that has been applied broadly across behavioral domains; from ethology (Bell et al., 2015; Bell and Pellis, 2011) and infant development (Plooij and van de Rijt-Plooij, 1990; Rijt-plooij and Plooij, 2013), to sociology (Mcclelland, 1996; McClelland, 1994, 2004) and psychotherapy (Carey et al., 2012; Higginson et al., 2011; Mansell et al., 2009). It has also been utilized as a foundational architecture for the development of autonomous robots (Young, 2017).

The PCT architecture has been tested most rigorously within the domain of motor control and learning. This is both because motor tasks are an example of real-time continuous control with highly measurable variables, and because the PCT model has explanatory value within motor control domain. For example, hierarchical PCT provides an elegant solution to the question of how voluntary movement is produced without the inhibition of postural control systems that would otherwise compensate such deviations in posture (Barter et al., 2015; Yin, 2014b, 2014a). In this scheme, voluntary movements are enacted by changing the reference value for the limb position control unit within the hierarchy. This has cascading effects, altering the reference values for the subordinate control units down to the postural control systems (velocity, body configuration, joint angle, muscle length, and muscle tension). In this way, the postural control systems drive the body to the intended position and maintain it (Yin, 2014a). Consequently, postural control is not disengaged and disturbances during the movement can be compensated on-line by the postural control systems. As stated, the PCT model has been tested most rigorously within simple control tasks such as manual tracking. This is because the tracking task is an environment in which control movements can be observed and measured while disturbances and feedback functions can be specified and manipulated.

The first PCT manual tracking study was published in 1978 (Powers, 1978). This article formalized the perceptual control model of behavior (Box 2) and laid the foundations for quantitatively analyzing an individual’s intentions by simulating their behaviour. Powers applied a known disturbance to the cursor during target tracking and measured the handle movements used to compensate the disturbance and control the cursor position. Powers was able both to quantitatively estimate the participant’s reference value, and to show that the logic of the PCT model could be leveraged to make inferences about which perceptual variable the participant was controlling. Powers argued that this approach could be taken to quantify intentions outside the lab. Since Powers’ 1978 article, many tracking experiments have been conducted. Several of these have utilized and extended Powers’ model. Hierarchical PCT models have been developed to account for simultaneous control of multiple perceptual variables and account for novel alterations to the feedback path (Marken, 1986, 1991). Several articles have attempted to determine individuals’ intentions (reference values) and quantify individual differences in tracking performance and parameters (Bourbon et al., 1990; Parker et al., 2017; Powers, 1989). While these tracking studies provide the primary empirical evidence base for PCT, and despite the broad application of the theory across domains, no review of the evidence for the theory provided by these tracking experiments, nor an appraisal of their methodological quality have been conducted to date.

In addition to PCT tracking studies, many tracking experiments have been conducted utilizing alternate models in recent years. These studies have been motivated by observations of inconsistencies between servo-models and human tracking. For example, the observation that tracking evolves as a series of short, intermittent movements (Abdel-Malek and Marmarelis, 1990; Gawthrop and Wang, 2011; Gollee et al., 2017; Inoue and Sakaguchi, 2014; Miall and Jackson, 2006; Miall et al., 1993), proposals concerning the mitigation of neural transmission noise and delays (Foulkes and Miall, 2000; Gauthier et al., 1988; Miall and Jackson, 2006; Stepp and Turvey, 2017; Vercher and Gauthier, 1992; Voss et al., 2007), observations of anticipatory tracking movements (Poulton, 1952b; Stepp, 2009; Stepp and Turvey, 2017; Voss et al., 2007), especially those made over short intervals of target occlusion (Fine et al., 2014; Mrotek and Soechting, 2007; Von Hofsten et al., 2007). Many of these studies are influenced by alternate motor theories, such as optimal control theory (Todorov, 2004) or active inference (Adams et al., 2013; Friston et al., 2011). Thus, a review of the evidence and limitations of the PCT model with respect to the tracking literature would be both relevant and useful. Additionally, a review of the evidence for PCT is timely given the broad application of the theory over the last decade and the frequent parallels and contrasts drawn between PCT and popular predictive coding accounts in recent years (Kumar and Srinivasan, 2014; Seth and Friston, 2016; Seth and Tsakiris, 2018).

The current article reports a systematic review of the evidence for PCT from tracking experiments. The first aim was to evaluate the proposal that manual tracking behavior is a process of perceptual control, and identify limitations to the PCT model of tracking. The second aim was, more broadly, to determine the state-of-the-evidence for the principles of perceptual control theory and its approach to studying human behavior. A summary of the findings from PCT model simulations of tracking is presented alongside an assessment of their methodological quality. Implications for tracking research and for PCT in action control are discussed.