In this article, we present a teaching sequence to introduce the topic area of neural networks with the help of Lego robots. Using the Lego system, we implement applications in the area of home automation in a large doll's house. With the help of neural networks, the home control system “learns” the behaviours of its residents and makes it individually configurable for each resident. The learning scenario enables differentiating tasks, which can be selected by the students themselves and thus allow all students to be successful. The lesson introduces the functionality of perceptrons and small multilayer neural networks on the basis of logical expressions.
This enables internal differentiation as well as a range of complexity with regard to
Smart Home solutions [
The target group with which this teaching scenario has been tested repeatedly
[
In this article, we want to expand the example of home control with physical
computing systems (here: Lego Mindstorms [
For our Smart Home application with neural networks, we use the Lego-NXT
Robot [
Our Smart Home application is an automatic window control system. This example should be viewed as representative for many other Smart Home applications. The window is to be opened and closed by means of a motor. On the one hand, this should be possible manually by using a touch sensor. The state of the window can be adjusted with the touch sensor. If the window is open, it can be closed by using the touch sensor. If the window is closed, it can be opened using the touch sensor.
Our neural network should be able to open and close the window automatically depending on the preference of the resident. To do this, it processes the input values of two sensors: a sound sensor and a light sensor. For simplicity’s sake, there are only two different input values for each sensor. The sound sensor can measure loud or quiet. The light sensor can measure light or dark. Both sensors are attached to the outside wall of the house.
In order to motivate the integration of a neural network, we consider the following example scenario.
Take the house of the Smith family: It has several bedrooms, in which the same new intelligent window closing system is to be installed. However, the residents of our house all have very different preferences.
Uncle Peter always sleeps with the window open, regardless of how noisy or quiet it is outside. But he always keeps the window closed during the day, because he is at work and a closed window seems more secure to him. Above all, he doesn’t want any insects to fly into his room during the day.
Mum Kate also sleeps with the window open, regardless of how noisy or quiet it is outside. During the day however, she opens the window even if it is loud outside, because that’s when the children play in the garden and Mum Kate wants to be able to hear them. The windows are only kept shut during the day when it is quiet outside.
Grandma Emmy has her very own peculiarities. She opens the window during the day when it’s noisy so she can hear her grandchildren playing and the birds singing. On quiet days, the window is kept closed. She likes her peace and quiet during the night. As such, she only opens the window at night when it is quiet outside. If it is noisy outside, however, she closes the window at night because she cannot sleep otherwise.
Dad Eric always works night shifts. For this reason, he keeps the window closed at night because he’s not at home. Dad Eric sleeps during the day. But he only opens the window during the day when it’s quiet outside. If not, the window is kept closed so he can sleep in peace. 3
We want to implement a small neural network, with which simple logical
operations can be realised. The network should have only binary inputs and outputs, but
have multiple layers, for example to also realise an XOR operation [
We only choose binary inputs and outputs so that the pupils can manually retrace the adjustment of weights in our small neural networks. In our learning scenario, we use the sound sensor and the light sensor. All sensors should yield binary values, the pupils can decide the thresholds themselves so that the sensors yield the values -1 or 1.
The following coding should apply:
• sound sensor S: S corresponds to S = 1 (loud), S̅ corresponds to S = -1 (quiet)
• light sensor L: L corresponds to L = 1 (light), L̅ corresponds to L = -1 (dark)
In this way, we can represent all logical AND operations that can be formed from
L and S (in the following referred to as minterms) as perceptrons [
If the sum of inputs multiplied by the corresponding weights in a neuron is larger
than or equal to the threshold of the neuron, then the neuron yields the output 1, in all
other cases -1. For the output neuron of the final layer we define:
• Output (action) = 1 => Window open
• Output (action) = -1 => Window closed
A perceptron with two input points, which is connected to the two sensors S and L,
can be implemented and also trained. We will not go into any more detail at this
point, as the implementation results directly from a reduction of the implementation
of a multi-layer network which we will describe in detail later, or which can be
looked up in [
From this, we can derive the necessity of using multi-layer networks, as Grandma Emmy’s configuration corresponds to the logical term NOT-XOR.
Our small neural network should be able to realise simple logical operations. We
continue to use two input variables and a total of three neurons, which we want to
connect as follows (see Fig. 2):
A few words on the use and legitimation of this type of network (cf. multiple adaptive
linear neuron [
The network in Figure 2 is thus sufficient for any “desired window opening” of our
residents. In didactic terms, however, we also have the advantage that the weights in
the final layer remain constant, and that during the learning phase, we only have to
consider changes to the weights in the first layer. Processes such as backpropagation
[
First, we translate the requirements of our residents into neural networks (whereby there can be several options). This is also a possible initial practice task for pupils, which can enable a reflective consideration of the training of our neural networks later on.
For the implementation, we separate our problem into three stages. First, we look at the mechanics. To do so, we equip a window with a motor and open and close it by using a touch sensor, which we test.
In the second stage, we implement the neural network, but set the weights by hand in the source code. In this way, we can realise and test the control for Grandma Emmy or Mum Kate or Uncle Peter.
In the third stage, the weights should adjust automatically and our Smart Home solution should be equally suitable for all residents. For this, the neural network should control the window motor. If, however, a resident is unsatisfied and subsequently adjusts the window manually using the touch sensor, the weights should change; thus, the neural network should “learn” the behaviour of our residents.
The “learning” of our neural network is comprised quite deliberately only of the change to the weights at the edges. The threshold should remain as unchanged as the number of neurons or the existence of links. For the pupils however, a systemisation stage of the topic should make clear that this represents a didactic reduction, and that a neural network can, in principle, be changed in other ways during the running time.
The code that implements all three stages is described in the following. 4.1
Explanation of the implementation First, we will explain the use of variable names in the following figure (Fig. 3):
The sound sensor yields the value S = 1 if it is loud outside, and S = -1 if it is quiet outside. This is encapsulated in the block: “get value sound sensor”
The light sensor yields the value L = 1 if it is light outside, and L = -1 if it is dark outside. This functionality is also encapsulated in the block: “get value light sensor”
The weights on the incoming edges of the upper neuron of the first layer are named ws1 and wl1. The neuron yields the result -1 or 1, which is stored in the variable ir1. (ir1 stands for “interim result 1”). The following applies: If the sum of inputs multiplied by the associated edge weights is larger than or equal to the threshold, then the output is 1 (the neuron fires), in all other cases it is -1.
In this case, mathematically stated: ir1 = {
1, if (S ∗ ws1 + L ∗ wl1) ≥ 2 −1, else ir2 = {
1, if (S ∗ ws2 + L ∗ wl2) ≥ 2 −1, else outputNeuron = {
1, if (ir1 ∗ 1 + ir2 ∗ 1) ≥ 0 −1, else
(3) (4) (5)
At the start of the program, all weights are set to zero and the light source of the light sensor is “switched off”.
As we want to be able to open and close our window manually using the touch sensor on the one hand, and automatically on the other hand, a further variable state is introduced. If state=0, this means that the window is currently closed. Variable state=1 means that the window is currently open. At the start of the program, the variable state is set to zero. The code for the automatic window control is shown in Figure 5.
Let us now consider the manual control. The window can be opened and closed manually with the touch sensor. If the touch sensor is activated and the window is open, it will be closed. If the touch sensor is activated and the window is closed, then it is opened. For this, we use our variable state. However, if the window needs to be opened or closed manually, this means that our neural network is not configured according to the requirements of the residents. The weights thus need to be adjusted.
If the window is closed and then opened manually, the weights have to be increased so that it would actually have been opened automatically.
If the window is open and then closed manually, the weights have to be decreased, because it should actually have closed automatically (Figure 6).
We carry out several runs with the same settings of input variables until the weights are set for this one case. This also decreases the number of training runs.
Finally, let us consider the blocks “lift the weight” and “reduce the weight” (see Fig. 7). The learning rate was set to 0.5 at the start of the program, and is stored in the variable “learning rate”.
If the window is open and is closed manually, then the weights have to be decreased. Remember: The window is open if the value of outputNeuron=1. It should have been -1, otherwise the window would not now have to be manually closed. If the value of outputNeuron=1, there are three possibilities: • ir1=1 and ir2=-1 • ir1=-1 and ir2=1 • ir1=1 and ir2=1
In the first two cases, the “faulty” neuron is identified in the intermediate layer (the one yielding the value of 1) and either of the weight pairs ws1/wl1 or ws2/wl2 have to be decreased in accordance with the input value S or L (if S or L is negative, the weights have to be increased, which occurs implicitly through the multiplication of S or L with the learning rate). In the third case, it is not clear which is the “faulty” neuron, which is why one of the two neurons are selected at random and the weights of the first layer of this neuron are changed.
The process for increasing the weights is similar.
If the window is closed and is opened manually, then the weights have to be increased. Remember: The window is closed if the value of outputNeuron = -1. It should have been 1, otherwise the window would not now have to be manually opened.
However, the value of outputNeuron can only be -1 if both neurons in the intermediate layer yield the value -1, that is, ir1 = ir2 = -1. The “faulty” neuron cannot be identified. Thus, one of the two neurons is selected at random. The weights of the first layer are increased for this neuron. (If S or L is negative, the weights have to be decreased, which occurs implicitly through the multiplication of S or L with the learning rate).
Our learning scenario describes an introduction to the functionality of artificial
neural networks. Our neural networks implement simple logical functions, in
accordance with [
Using a motivating example scenario with everyday relevance and the opportunity for a haptic, enactive experience with technical systems, the pupils learn the following: • The pupils can describe and implement the functionality of a perceptron • The pupils can name the limitations of a perceptron • The pupils can describe and implement the functionality of a small, multi-layer network • They recognise that the implementation of neural networks represents a deterministic algorithm, which they are familiar with • The pupils can understand and reflects on the adjustment of weights of a neural network through manual feedback • They recognise that due to a training phase, particular parameters of the algorithm during the runtime are adjusted, and the algorithm can thus be configured differently • They recognise that the parameter adjustment (depending on feedback) can be different, and therefore future results of the algorithm are in accordance with the parameter settings • They can identify a previously unfamiliar problem-solving strategy, which is that one and the same algorithm is used for different applications, in that it is trained by the adjusting of parameters and thus configured • The pupils recognise that the algorithm does not achieve a level of consciousness or similar, such as might be assumed for “artificial intelligence”. In this article, there is no survey of learning groups before and after the lesson to confirm the learning aims. This is due to the current Corona crisis and will be carried out as soon as possible. 6
Our learning scenario is intended to offer an introduction to the basic concepts of neural networks. As such, it aims to show that the “learning” of neural networks is equivalent, among other things, to the adjustment of particular parameters (here: the weights) during the runtime on the basis of feedback.
The author views the benefits of physical computing to include an enactive dealing
with informatics and the creation of operable, impressive products, even if these are
not particularly complex in informatics terms. In this way, even less able pupils can
be successful. At the same time, the Smart Home offers a learning environment that is
motivating and relatable to the lives of the pupils. With many small functions
(automatic window opening, automatic ventilation control, …) that can be integrated into a
Smart Home, pupils can choose their own tasks, which increases their pride in the
later product [