A robot that can learn from its environment is a dream of many.
It is possible to create a robot that learns from its surroundings in a way that is as natural as possible.
In this article we explore how one such machine is made, and how it might be adapted for use in the future.
The basic idea is that the robot learns to recognise patterns in its environment, and then to adjust its behaviour accordingly.
It then repeats the process until it has learnt to use the environment correctly.
To make the process as natural and simple as possible, robots often have built-in learning systems.
These are often referred to as ‘simplifying algorithms’.
This is because they do not take into account the way humans learn.
Instead, they use ‘contextual learning’ to learn to recognise and adapt to the environment.
A typical ‘context-based’ learning system is described in the paper ‘Learning from the environment: a model of human learning from the context of the environment’.
A context-based learning system consists of a set of rules that can be learnt by the robot from its previous behaviour.
The rules are often expressed as ‘learned rules’ that are passed on to the next time the robot is tested in the same way.
Examples of learning systems are the learning algorithms shown in the figure below.
In general, learning from a particular context is better than learning from its absence.
For example, the learning system shown in figure 1 would have learnt that the mouse should move towards the red box when it is pressed, even if the box was hidden.
However, the system would have been able to predict that if the mouse was pressed and the box were hidden, the box would move towards it.
This would have helped it to adapt its behaviour if it had been pressed and then hidden.
Learning from the absence of a particular element of a stimulus can be particularly useful.
In the figure shown above, we see that the system is able to learn that a particular area of the screen is not illuminated, even though it is illuminated.
The system then uses this knowledge to modify its behaviour to avoid using that area.
This is illustrated by the ‘silly mouse’ example below.
Figure 2 shows a simplified version of the ‘mouse’ example.
The ‘sly mouse’ is a simple example of a context- based learning system, but the principles can be applied to any system that is able and willing to learn from the behaviour of other systems.
Examples include systems that learn by playing video games, learning how to drive cars, or learning how the brain responds to sound.
Figure 3 shows an example of the learning from an audio stimulus that we have previously shown to be a powerful example of context-sensitive learning.
A sound stimulus is presented in a room and the robot reacts by moving around the room, using the sounds to find its way.
Figure 4 shows a similar example of an audio system.
The sound system shows how to recognise how an animal responds to a particular sound by using an audio cue.
This example demonstrates how learning from other systems can be used to design more complex systems.
The most common examples of context sensitive learning are learning algorithms that learn to detect patterns in an image, or to detect changes in the shape of an object when it changes in size.
The images shown in figures 1 and 3 are examples of images that we use to create our ‘contexts’.
In general a new image is formed when the object or image is changed by the movement of a human or animal.
The human or robot that is moving then takes the shape that is most appropriate for the image.
The algorithm then adjusts its behaviour by detecting changes in this shape.
A context sensitive algorithm that is capable of learning from images of objects is known as a ‘context sensitive learning algorithm’.
In a context sensitive training system, the shape is determined by the image being changed.
If the shape changes then the image must be changed too.
The key point is that these are ‘trained’ images that are only ‘learnable’ if they are shown to the human or other animal.
These images are then used to create the next image.
In order to learn the shape from an image that is only visible in the new context, the algorithm has to learn a ‘learn-by-example’ strategy.
The model that the algorithm uses to solve the problem is known by the term ‘learning algorithm’.
An example of this learning algorithm is shown in Figure 5.
Figure 5 shows an image of a car in the scene shown in Figures 1 and 2.
The robot that has just learned to drive the car, the ‘blue car’, has learned to look for the ‘red car’ when it moves towards the left of the camera.
This makes sense as the robot has learned the ‘learn to move left by looking for the red car’ behaviour.
However the robot also has to solve another problem that is not solved by the car.
When the robot looks for