r0mai / car-game Goto Github PK

If I train the network on say track A, and then start working on a new track called B, I would like to be able to keep the results from track A. A configurable output file seems ideal, but as a default, perhaps instead of best.car the results could be stored in best.car.TRACKNAME

Cheers.

I would like to be able to limit the time needed for training the AI, by way of a command line switch such as:

--run-time=DURATION
--generation-count=NUM
--fitness=NUM

Either of these would be okay, just a mechanism to be able to batch run training jobs without needing an interactive ^C.

Cheers.

When there are multiple cars on each track, display info about each car as a text near the car.

Ability to create different shapes of tracks, not just simple circular maps.

Currently only single threaded mode works because of this. Possible solutions:

1. Lock Lua class.
2. Move fitness calculation to main thread.
3. Use one lua interpreter on each thread.

Refactor to use less memory.

Drawing functions should not belong to the class Model. It is a bad separation of responsibility. Drawing functions should all got to RealTimeGameManager, or any other classes instantiated only by RealTimeGameManager.

With more checkpoints, the AI has more information about the upcoming turn, so it can make more advanced manouvers.

Not yet fully reproduced, but it is suspected that it comes up only when the debug output is switched on for fitness calculation, i.e. when the calculator function returns with not only a number but a number and a string.

In interactive mode, draw a line to display where the car went during its route.

Calculating geometry on GPU would be much faster.

Work in progress

If I have a track which is say 1000 units wide and 1000 units tall, the window should be able to zoom out to fit the entirety of the track on the frame. Currently it seems to me, that the window zoom is fixed.

An alternative to making the window scale, is to scale the track itself, however distances may not be accurate if we scale the track down.

Cheers.

Work in progress

In interactive mode, change the color of the car gradually to indicate how much time it spent intersecting the track edge. Also display it in text.

Say I defined a track with points:

0, 0;
1, 0;
1, 1;
0, 1.

In this case, the circumference is 4, however this may be modeled after a real-life track, for which the circumference is known to be (say) 40. In this case we need to scale the points in the track by a factor of Sqrt[40/4].

This would be helpful in creation of real-life tracks.

Cheers.

Turn on/off

• rays
• checkpoints
• track boundary
• telemetry graphs (done)
• telemetry text
• car

Wori in progress

The following library could be used for smoothening the corners:
http://www.angusj.com/delphi/clipper.php

I would like to be able to run the ai training for a limited time, interrupt the training, then be able to continue from where it left off. This would prove useful to be able to create snapshots for the AI, or for continuing training on a different track.

For instance if the first round of training runs until generation X, then the training shouldn't necessarily start from scratch.

Cheers.