Reward Pays the Cost of Noise Reduction in Motor and Cognitive Control

Abstract

Speed-accuracy trade-off is an intensively studied law governing almost all behavioral tasks across species. Here we show that motivation by reward breaks this law, by simultaneously invigorating movement and improving response precision. We devised a model to explain this paradoxical effect of reward by considering a new factor: the cost of control. Exerting control to improve response precision might itself come at a cost--a cost to attenuate a proportion of intrinsic neural noise. Applying a noise-reduction cost to optimal motor control predicted that reward can increase both velocity and accuracy. Similarly, application to decision-making predicted that reward reduces reaction times and errors in cognitive control. We used a novel saccadic distraction task to quantify the speed and accuracy of both movements and decisions under varying reward. Both faster speeds and smaller errors were observed with higher incentives, with the results best fitted by a model including a precision cost. Recent theories consider dopamine to be a key neuromodulator in mediating motivational effects of reward. We therefore examined how Parkinson's disease (PD), a condition associated with dopamine depletion, alters the effects of reward. Individuals with PD showed reduced reward sensitivity in their speed and accuracy, consistent in our model with higher noise-control costs. Including a cost of control over noise explains how reward may allow apparent performance limits to be surpassed. On this view, the pattern of reduced reward sensitivity in PD patients can specifically be accounted for by a higher cost for controlling noise.

Keywords: decision-making; dopamine; drift-diffusion model; motivation; speed-accuracy trade-off.

Graphical abstract

Figure 1

Breaking the Speed-Accuracy Trade-Off A conventional account of the speed-accuracy trade-off is shown on top (A and B). Lower panels illustrate the inclusion of a precision cost (C and D). (A) As movement time decreases (and speed increases), accuracy declines. This is because faster movements require larger forces, which are susceptible to proportionally higher noise. Because noise is taken to be rigidly proportional to motor command size, behavior is constrained such that accuracy depends on speed. This is captured by a speed-accuracy trade-off contour (illustrated in blue in the schematic). (B) Standard motor control models determine the optimal movement as the one that gives the highest average payoff. The expected value (EV) of a movement depends on the motor command u. EV can be expressed in terms of the reward, discounted by the movement time (pink). Larger motor commands give faster movements, leading to earlier reward, which is more valuable (here we use hyperbolic temporal discounting). The reward is further reduced by movement error, which reduces the probability of success (green). Under the orthodox view, this error is determined by noise proportional to the motor command. This means that faster movements will have less accurate endpoints and thus have a lower chance of winning a reward. The final term is the energetic cost of the motor command itself (blue). The balance between time and accuracy is governed only by the movement speed, which guarantees that speed and accuracy trade off with one another. (C) We suggest that reward has a motivating effect that permits both faster and more precise behavior. Incentivization by reward thus produces a change in a direction perpendicular to the blue line (red arrows), contrary to the speed-accuracy trade-off. By allowing both force and movement precision to be varied, our model allows two degrees of freedom over speed-accuracy space. (D) In order to explain violations of the speed-accuracy constraint, we introduced an additional precision command. The precision command reduces noise (u_P, shown in red), complementing the usual force command (u_F, blue). We propose that this command is itself costly, in the same way as the force command u_F, leading to a cost term |u_P|² + |u_F|². Optimizing EV by selecting both the precision and force would allow accuracy to improve independently of speed, but constrained by this cost. Higher incentives allow a greater investment in precision, rather than a trade-off with speed, so genuine performance improvements are possible.

Figure 2

The Costs of Inaccuracy, Slowness, and Control Our model incorporates three costs: inaccuracy is expensive because errors are not rewarded, slowness is expensive because a reward is less valuable when delayed (temporal discounting), and we further suggest that control over errors is itself expensive. Therefore, a three-way balance obtains. (A) The vertical axis represents the subjective value of a given movement. For a given reward R, temporal discounting causes the reward’s subjective value to fall as movement times get longer (dotted lines). However, moving faster entails greater energy expenditure (dashed line; negative value implies a cost). The net value (solid lines) is the sum of these two components, showing that the optimal movement is faster with higher reward [36]. (B) The probability of winning a reward, P_win, could depend on the endpoint of the movement being accurate. The cost of precision allows the endpoint variability to be reduced at a cost. The probability of landing on a fixed-size target can be increased if a “precision cost” is paid (dashed line). Precision increases the average gain from winning (dotted line), as shown for three different reward levels. The net value (solid lines) illustrates that the optimal movement is more precise with increasing reward. (C) If both speed and accuracy are both free to vary, the optimum pair can be determined as a function of reward. Reward increases the optimal movement speed and, when temporal discounting is not too large, reduces the optimal endpoint variability.

Figure 3

Optimal Control Model to Explain the Effect of Reward Incentives In order to account for the ability of reward to improve both speed and accuracy, we hypothesized that in addition to a “vigor” or force signal (u_F) that determines a movement’s speed, individuals are also able to select a “precision” signal (u_P) that determines the amount of variability in a movement. Crucially, this precision signal is also costly. (A) Each given motor command, i.e., a pair of force and precision u = (u_F, u_P), has an EV. The image shows EV as a function of u, with the best combination as blue and worst as red. The value depends on three effects. First, the reward available is temporally discounted by the time taken by the movement, e.g., by hyperbolic discounting $1/\left(1+k/\sqrt{u_F}\right)$. Second, this reward is only obtained if the movement is on target. We assume a Gaussian variation Φ of the endpoint proportional to the size of the motor command. Third, although we can go faster to reduce temporal discounting (increasing u_F) and be more precise to reduce error (increasing u_P), both of these incur a cost proportional to the squared control signal, u². This leads to an optimal combination of force and precision for each movement, u^∗. (B) The optimal motor command for a situation depends on the reward level R and on two subject-specific parameters: the discount rate k and the noise-control cost σ. The optimal precision (upper panels) and force (lower panels) both increase with increasing reward (y axis), indicating that reward induces greater “spending” on both speed and accuracy. However, precision and force are differentially influenced by reward, and the balance depends on the urgency (temporal discount, k, left panels) and error constraints (encapsulated by σ, right panels). (C) The optimal commands determine the velocity and duration of each movement and the amount of variability for a desired movement amplitude. Reward always increases velocity (lower panels). However, variability may increase or decrease with reward (upper panels), depending on σ and k. A subject with minimal discounting (e.g., k < 0.5) becomes less variable with higher reward, whereas a subject with high discount rates (e.g., k > 1) in fact tends to become more variable with higher reward (upper panels) as they are under greater time pressure, i.e., trading speed for accuracy. These effects are re-plotted on different axes in Figure S5.

Figure 4

Cost of Control Applied to Drift Diffusion (A) The drift-diffusion model assumes an accumulator integrating incoming information at a fixed drift rate (μ), subject to noise (σ), until it reaches a threshold (θ). The red line illustrates the trajectory in an example trial. Blue histograms indicate the distribution of response times for correct and incorrect responses. Increasing the threshold leads to more accurate decisions, at the cost of slower responses. In order to account for violations of the speed-accuracy trade-off, we introduced a costly noise-reduction parameter (u_P), similar to our extended motor control model. This permits the optimal combination of threshold and precision to be chosen. (B–E) Simulations provide reaction times and accuracy (i.e., when the decision terminates, and whether it is at the positive or negative boundary) for a variety of signal sizes (μ), noise (σ), and reward levels (R). For each condition, the optimal pairing of threshold (θ) and precision (u_P) is selected to maximize value (EV). The value of a pair was calculated as accuracy multiplied by reward, temporally discounted by the reaction time. (B and C) As reward increases, it is optimal to increase the precision and lower the decision threshold. (D) This leads to improved accuracy with reward. (E) When the signal-to-noise ratio is high, reward encourages faster responding; however, when the decision is noisy, reaction times actually increase with reward, despite falling thresholds—producing a speed-accuracy trade-off.

Figure 5

Oculomotor Capture Task with Trial-wise Incentives (A) Three equidistant discs were dimly illuminated. At the start of each trial, participants had to fixate one disc, which was brightened. A recorded voice gave an auditory reward cue of “0p maximum,” “10p maximum,” or “50p maximum,” which indicated the maximum amount of money that could be won if participants were fast to look at the target on that trial. After a variable foreperiod, the other two discs were illuminated asynchronously, with a delay of 40 to 120 ms. Participants were instructed to look as fast as possible to the second disc. Thus, the first onset acted as an early onset distractor, and the second disc indicated the target. (B) After gaze arrived at the target, participants were rewarded according to reaction time. Reward was calculated as a fraction of the maximum available, using an exponential falloff. The falloff was determined adaptively using quantiles of the last 20 trials, in order to maintain the difficulty level over the course of the experiment. (C) On approximately 30% of trials, gaze was “captured” by the distractor (errors), resulting in a brief saccade to the first disc, followed by a corrective saccade to the target. The trajectory of gaze was classified according to whether the first saccade terminated on the target or on the distractor. Correct trials exhibited a variety of curvatures; each trial is colored according to the initial direction of the eye velocity. (D) The task provided four measures of performance. Both speed and accuracy could be examined for motor execution of the saccade and for selection of the correct target.

Figure 6

Effects of Reward on Saccades in Healthy Participants (A) For correct trials, the mean peak velocity of saccades increased with higher incentives, demonstrating invigoration by reward. Error bars indicate within-subject standard error. (B) Endpoint variability (standard deviation of the saccade amplitudes) in each condition became less variable with increasing incentives, indicating that reward can improve motor precision. (C) For correct trials, RTs were faster for higher incentives. (D) The rate of oculomotor capture (proportion of trials on which the first saccade after the onset was directed to the distractor, i.e., error trials) was reduced with increased incentives, indexing improved accuracy. Error timings are shown in Figure S2 and times to correct errors in Figure S3. (E) Plotting the data from Figure 5 as accuracy versus RT (where accuracy is defined as percentage of responses that were directed to the target and not to the distractor) demonstrates how, with increasing incentives, reaction time decreased and accuracy simultaneously improved. The inset shows how this relates to Figure 1C: the speed-accuracy trade-off is broken. (F) Conditional accuracy plot shows how, for a fixed reward level, accuracy improved with increasing RT, but this relationship was shifted by incentives, with the greatest differences evident at short RTs. The gradient of each curve is always positive, indicating that for a trials within a single reward level (i.e., constant incentive), the speed-accuracy trade-off held. The plot shows the proportion of saccades that went to the target, in a sliding window along the RT distribution, width 20% quantiles. Patient data and model are shown in Figure S1.

Figure 7

Reduced Reward Sensitivity in Patients with Parkinson’s Disease (A) PD patients had decreased reward sensitivity, as demonstrated by a shallower slope. This is consistent with impaired invigoration by reward. Overall velocities were also marginally slower. (C) Saccadic amplitude variability was not significantly abnormal in PD. (B) Reaction times were slower in PD and showed reduced reward sensitivity. (D) Patients showed weaker effects of reward on improving distractibility, as measured by oculomotor capture (i.e., they did not reduce their error rate in response to incentive), compared to controls. Between-subject correlations are shown in Figure S4.