![]() ![]() The Q-learning algorithm picks in each state the action which promisses the best result. In case the car crosses the goal line the reward is Qgoal. In case the car leaves the track (crashes) the reward is value Qfail. Nevertheless it is present and ready for experimentation. In this implementation an alternative method to request exploration is present (controlled via epsilon, see below) and this might mean Q_0 is not needed. Large values will encourage more exploration. The default in 0.0 but larger values can lead to interesting results. It value must be larger than Qfail (see below). The Q_0 value is the initial value for all inputs to the Q-function. The value must also be between 0.0 and 1.0 and again it better not be 0.0. This is the discount factor which determines how much weight future results are given. The default is 0.1 but smaller or larger values can and should be attempted. It better should not be zero since this means no learning at all. This is the learning rate which determines how quickly the results of next steps influence the prior step. The parameters are part of the Q-function algorithm: The program provides for each of the parameters which can be used to influence the learning algorithm default values which can be seen when python3 race.ph -help is executed. If only the edge of the track is given (as in the picture below) it is important that the edge line is a continuous line, fully horizontally or vertically connected. It is not necessary to fill the outside of the track in this color but it could be done. (0,0,0) which is black is the background where there is no goal line, starting point, first step, nor the edge of the trackĪny other value corresponds to the edge of the track and/or the outside of the track.(0,0,255) which is blue is the target of the first step and therefore defines the initial speed.(255,0,0) which is red corresponds to the starting point which must be on the start line. ![]() (255,255,0) which is yellow corresponds to the start/goal line.Support for other types of image files could be added but is not present.įive different colors are used with the following RGB values: I.e., the file must contain a palette which is indexes by the byte-sized values which make up the actual picture. The file format is expected to be indexed 8-bit color PNG. The track should be a picture of the track. The various parameters specified as parameters.A PNG file with the picture of the track.The inputs for the program come in two forms: Some non-standard Python modules are needed but they are all available in the Fedora repository and can be installed with dnf. It is not long and should therefore be understandable to the interested reader. The program is implemented as a Python script for easy understanding. It is therefore important to know about as many techniques as possible. What people forget is that many or even most solutions using neural networks also use other techniques in combination or series which can significantly increase the efficiency of the model. The problem is also chosen as one which work well with non-NN solutions, algorithms which are often drowned out in today's world focussed on neural networks. ![]() This is an example for a solution of a problem which might be prohibitively expensive to solve using non-probabilistic methods. ![]()
0 Comments
Leave a Reply. |
Details
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |