Lieder and Griffiths demonstrate a capacity of the principle of optimal use of limited computational resources (resource-rationality principle) to account for a wide variety of observations in multiple disciplines, including psychology, neuroscience, linguistics, and economics. However, they have overlooked recent developments in the field of neural control of human goal-directed movements where the explanatory power of that principle has been demonstrated. We briefly review those developments below and show how several different pieces of evidence support the resource-rationality principle.
The first step toward making a connection between cognitive resources and characteristics of human motor performance was made yet by Fitts (Reference Fitts1954). He demonstrated that if experiment participants are asked to move to a target, the movement time is linearly proportional to the index of task difficulty that Fitts described as the amount of information needed to be processed to achieve required precision. When the distance to the target is fixed, difficulty in task is determined by the size of the target, and the smaller the size the longer the movement time, which is known as speed-accuracy tradeoff. This relationship shows that the neural system controlling movements can tailor the precision of information processing to the required precision of motor task performance and movement speed to precision demands. The phenomenon of speed-accuracy tradeoff has a simple interpretation. If the movement is too fast, there may not be enough time for accumulating the minimal amount of information required for sufficiently precise state estimation and decision making.
Later studies have corrected the concept formulated by Fitts and showed that the movement control system uses a two-phase strategy. In particular, Shimansky and Rand (Reference Shimansky and Rand2013) demonstrated that during the initial phase, the use of neural resources for processing sensory information is minimized, while the control system relies on the internal model of the controlled object's dynamics. Speed-accuracy tradeoff is violated in that phase. The final phase is performed with speed-accuracy tradeoff, with the precision of sensory information processing being determined by the required movement precision at the target.
To account for their findings, Shimansky and Rand (Reference Shimansky and Rand2013) suggested that, since neural computations involved in information processing are costly, the cost of the “neural effort” should be minimized whenever possible during performance of motor tasks. Using the optimality approach, they formally introduced the “neurocomputational” cost (the cost of neural effort, or the cost of cognitive resources in terms of Lieder and Griffiths) as a vital component of the criterion (called “utility function” by Lieder and Griffiths) determining movement control optimality. Thus, the concept of two-phase strategy can be viewed as a generalization of the resource-rationality principle to reaching movement control.
The notion of the neurocomputational cost was further used to account for a hierarchical organization of control of the limb's joints that is typically observed during human movements (Dounskaia and Shimansky Reference Dounskaia and Shimansky2016). Namely, different joints of the limb (e.g., the shoulder and elbow) usually play different roles in movement production. One (“leading”) joint is rotated actively by the muscles spanning the joint, while the other joint “trails” due to passive factors, including gravitational torque and “interaction torque” caused by motion of the leading joint. This “trailing joint control pattern” is analogous to cracking a whip by swinging its handle, although the trailing joint musculature can interfere and adjust motion of this joint to task requirements.
Multiple studies have demonstrated the trailing joint control pattern during various types of arm movements (for review, see Dounskaia Reference Dounskaia2005; Reference Dounskaia2010). As discussed by Dounskaia and Shimansky (Reference Dounskaia and Shimansky2016), this pattern is a result of movement optimization, which is apparent from a tendency to maximally exploit passive torques for rotation of the trailing joint and from an observation that the contribution of passive torques to control of the trailing joint increases with development of skill. Dounskaia and Shimansky (Reference Dounskaia and Shimansky2016) used the information theory to show that the trailing pattern decreases the neurocomputational cost by reducing the amount of information that needs to be processed for joint coordination. Indeed, active control and coordination of all joints requires estimation of joint positions and development of corrective control commands at each moment of time. The trailing control pattern allows delegation of joint coordination mainly to passive torques and spinal reflexes to reduce the need for expensive neurocomputational processing of external sensory and proprioceptive information at the cerebral cortical level.
The tendency to reduce the neurocomputational cost has a strong potential to account for many other motor control phenomena. For example, causes for differential stability of various multi-limb coordination patterns, including bimanual movements, remain an object of debates (Swinnen Reference Swinnen2002). A comparison of different coordination patterns in terms of cognitive resources required for state estimation and decision making during generation of corrective control commands to each limb is a promising approach to account for experimentally observed differences in pattern stability. Theories that suggest simplification of control, for example, through the use of muscle synergies and motor primitives, and through reducing movement variability relevant for the task and ignoring irrelevant variability (Bruton and O'Dwyer Reference Bruton and O'Dwyer2018; Giszter Reference Giszter2015; Scholz and Schoner Reference Scholz and Schoner1999) implicitly represent the tendency to minimize this cost.
Finally, an application of the principle of the neurocomputational cost minimization to human movement control suggests that a learning process contributes to emergence of strategies that minimize the use of cognitive resources. This hypothesis is supported, for example, by an observation that more skillful movement performance is associated with the use of a more pronounced trailing joint control pattern, that is, the more intensive use of passive torques and spinal neural circuitries for production of training joint motion (Dounskaia and Shimansky Reference Dounskaia and Shimansky2016).
In conclusion, the extension of the resource-rationality principle to the field of human movement control described here increases the generality of this principle. This generalization will help to advance the principle of resource-rationality in both cognitive and motor control research fields.
Lieder and Griffiths demonstrate a capacity of the principle of optimal use of limited computational resources (resource-rationality principle) to account for a wide variety of observations in multiple disciplines, including psychology, neuroscience, linguistics, and economics. However, they have overlooked recent developments in the field of neural control of human goal-directed movements where the explanatory power of that principle has been demonstrated. We briefly review those developments below and show how several different pieces of evidence support the resource-rationality principle.
The first step toward making a connection between cognitive resources and characteristics of human motor performance was made yet by Fitts (Reference Fitts1954). He demonstrated that if experiment participants are asked to move to a target, the movement time is linearly proportional to the index of task difficulty that Fitts described as the amount of information needed to be processed to achieve required precision. When the distance to the target is fixed, difficulty in task is determined by the size of the target, and the smaller the size the longer the movement time, which is known as speed-accuracy tradeoff. This relationship shows that the neural system controlling movements can tailor the precision of information processing to the required precision of motor task performance and movement speed to precision demands. The phenomenon of speed-accuracy tradeoff has a simple interpretation. If the movement is too fast, there may not be enough time for accumulating the minimal amount of information required for sufficiently precise state estimation and decision making.
Later studies have corrected the concept formulated by Fitts and showed that the movement control system uses a two-phase strategy. In particular, Shimansky and Rand (Reference Shimansky and Rand2013) demonstrated that during the initial phase, the use of neural resources for processing sensory information is minimized, while the control system relies on the internal model of the controlled object's dynamics. Speed-accuracy tradeoff is violated in that phase. The final phase is performed with speed-accuracy tradeoff, with the precision of sensory information processing being determined by the required movement precision at the target.
To account for their findings, Shimansky and Rand (Reference Shimansky and Rand2013) suggested that, since neural computations involved in information processing are costly, the cost of the “neural effort” should be minimized whenever possible during performance of motor tasks. Using the optimality approach, they formally introduced the “neurocomputational” cost (the cost of neural effort, or the cost of cognitive resources in terms of Lieder and Griffiths) as a vital component of the criterion (called “utility function” by Lieder and Griffiths) determining movement control optimality. Thus, the concept of two-phase strategy can be viewed as a generalization of the resource-rationality principle to reaching movement control.
The notion of the neurocomputational cost was further used to account for a hierarchical organization of control of the limb's joints that is typically observed during human movements (Dounskaia and Shimansky Reference Dounskaia and Shimansky2016). Namely, different joints of the limb (e.g., the shoulder and elbow) usually play different roles in movement production. One (“leading”) joint is rotated actively by the muscles spanning the joint, while the other joint “trails” due to passive factors, including gravitational torque and “interaction torque” caused by motion of the leading joint. This “trailing joint control pattern” is analogous to cracking a whip by swinging its handle, although the trailing joint musculature can interfere and adjust motion of this joint to task requirements.
Multiple studies have demonstrated the trailing joint control pattern during various types of arm movements (for review, see Dounskaia Reference Dounskaia2005; Reference Dounskaia2010). As discussed by Dounskaia and Shimansky (Reference Dounskaia and Shimansky2016), this pattern is a result of movement optimization, which is apparent from a tendency to maximally exploit passive torques for rotation of the trailing joint and from an observation that the contribution of passive torques to control of the trailing joint increases with development of skill. Dounskaia and Shimansky (Reference Dounskaia and Shimansky2016) used the information theory to show that the trailing pattern decreases the neurocomputational cost by reducing the amount of information that needs to be processed for joint coordination. Indeed, active control and coordination of all joints requires estimation of joint positions and development of corrective control commands at each moment of time. The trailing control pattern allows delegation of joint coordination mainly to passive torques and spinal reflexes to reduce the need for expensive neurocomputational processing of external sensory and proprioceptive information at the cerebral cortical level.
The tendency to reduce the neurocomputational cost has a strong potential to account for many other motor control phenomena. For example, causes for differential stability of various multi-limb coordination patterns, including bimanual movements, remain an object of debates (Swinnen Reference Swinnen2002). A comparison of different coordination patterns in terms of cognitive resources required for state estimation and decision making during generation of corrective control commands to each limb is a promising approach to account for experimentally observed differences in pattern stability. Theories that suggest simplification of control, for example, through the use of muscle synergies and motor primitives, and through reducing movement variability relevant for the task and ignoring irrelevant variability (Bruton and O'Dwyer Reference Bruton and O'Dwyer2018; Giszter Reference Giszter2015; Scholz and Schoner Reference Scholz and Schoner1999) implicitly represent the tendency to minimize this cost.
Finally, an application of the principle of the neurocomputational cost minimization to human movement control suggests that a learning process contributes to emergence of strategies that minimize the use of cognitive resources. This hypothesis is supported, for example, by an observation that more skillful movement performance is associated with the use of a more pronounced trailing joint control pattern, that is, the more intensive use of passive torques and spinal neural circuitries for production of training joint motion (Dounskaia and Shimansky Reference Dounskaia and Shimansky2016).
In conclusion, the extension of the resource-rationality principle to the field of human movement control described here increases the generality of this principle. This generalization will help to advance the principle of resource-rationality in both cognitive and motor control research fields.