Almost all human actions are sequential actions, in which an activity can be split up in different smaller sub-actions. Cooking, for example, can be split up in different smaller actions such as boiling the water or slicing up ingredients.
Since sequential actions are, thus, such an important part of human behavior, it has been studied a lot in different disciplines, such as psychology (Behmer, 1992), biomedical sciences (Rojas et al. 2008), and artificial intelligence (De Kleijn, Kachergis, and Hommel, 2014). An important paradigm used in sequential action is the Serial Reaction Time task (SRT). This task was developed 30 years ago and was mainly used in understanding the processes underlying in many human behaviors, such as learning (Nissen and Bullemer, 1987). In this task four positions are arranged horizontally on a computer screen.
At each of these positions a visual cue can appear. When such a visual cue appears on one of the positions, participants have to press a button, corresponding to that position. The visual cues are shown in a repeated sequence and the reaction time (RT) it takes for the participant to press the right button in measured. After a sequence trial, the participant are presented a random trial, in which it is not possible to predict the next stimulus. The SRT is, thus, a reaction time task in which participants learn and predict a sequence. This makes it a useful paradigm to study sequential learning.
Kachergis, Berends, De Kleijn, and Hommel slightly adapted to original SRT task (2014). In this task participants did not have to respond to the stimulus changes by pressing a button, but by moving the cursor to the corresponding stimulus. Therefore, the four stimuli were mapped to a square: one in the top-left (1), one in the top-right (2), one in the bottom-left (3), and one in the bottom-right of the square (4). In a training block the participants had to learn a sequence of ten stimuli: 4-2-3-1-3-2-4-32-1. In this training block they had to follow this sequence by moving the cursor to the corresponding square on the screen. After the training block the participant had perform a 2 generating task, in which they had to predict the next stimuli instead of following the sequence.
In their study they were able to replicate the results by Nissen and Bullemer (1987) and, thereby, measure extra context effects, such as the trajectory effect of the cursor movements (Kachergis et al., 2014, De Kleijn et al., 2016). Measuring these trajectory effects during the SRT give more insight in the predictive movements during the experiment. We can, for example, see if the participant already moves the cursor to the next stimulus location, by predicting the next location, so before appearance of the stimuli. This is an example of optimization behavior in sequential action.
These types of optimization behaviors have been extensively been studied, for example by Dale, Duran, and Morehead (2012). They performed a similar sequential action task to the one described by Kachergis et al. (2014), but with different sequence complexities. They found that with simple sequences, when people were more likely to be aware of the sequence, participant made larger predictive movements towards the next location. On the other hand, when the participants were not aware of the sequence and, thus, were not able to make a correct next prediction, they showed centering behavior.
In these cases they moved their cursor to the middle of the screen, so each stimuli location would be at the same distance. De Kleijn, Kachergis, and Hommel (2017) implemented the SRT task by Kachergis et al. (2014) in a virtual robot hand, which was controlled by an artificial neural network.
They used a two-layer feedforward neural network with a neuroevolution algorithm to optimize the network weights. The networks consisted out of 6 input neurons, 2 predictive neurons and 4 sensory neurons, 8 hidden neurons, and 2 motor output neurons, which controlled the robot arm. As input for the predictive nodes they used three conditions: an accurate prediction, no prediction, and random prediction.
The input in the accurate prediction was the location of the next cue, in the random prediction the input was one of the four locations, randomly picked, and in the no prediction condition there was an constant input of 0.0,0.0. The results showed that, in the random prediction and no prediction condition, the network developed the same centering behavior as the earlier studies in human participants.
Thereby, the network with the accurate prediction gained the highest fitness score, but that it evolved somewhat slower than the network with no prediction. There was also a difference between the no prediction and the random prediction condition. The network with the random prediction evolved slowest, and, thereby, scored the lowest fitness score. However, the difference between the random prediction condition and the no prediction condition was not significant.
Although the results were not significant, these results imply that the networks had trouble ignoring the random and uninformative input provided. It would be interesting to see if the same thing happens in human learning. However, no research has yet been done to human performance under these different circumstances. Therefore, in this study will be researched how humans perform when they know that they are unware of a sequence and how humans perform when they try to predict the sequence, even though this is not possible. The aim of this research is to see if humans have the same trouble ignoring the uninformative input provided by the previous trials as the neural network in the research by De Kleijn et al. (2017), this way, gaining more knowledge into human sequential learning. Therefore, the following research question has been formulated: does knowing that a prediction input is random stop the interference between the uninformative input and the performance on an SRT task and how does this influence the centering behavior? Based on the results by De Kleijn et al.
(2017) it is hypothesized that humans, as the artificial neural networks, have the same trouble ignoring the random prediction stimuli. However, knowing that they are see a random sequence will stop this interference. Therefore the RTs of the groups that 3 knows that the sequence is random will be higher than the RTs of the group that is unaware of the fact that they are learning a random sequence. Second, it is hypothesized that the groups that is aware of the random sequence will show more centering behavior than the group that is unaware of the random sequenc