marcelja / doodle-jump Goto Github PK

Let the player play multiple different games within the same generation.

• add other platform types
• increase number of platforms

We should alter the fitness function for that

Currently, networks are never deleted, taking up a lot of memory quickly.

Generation
Score

Otherwise a single instance that jumps on one platform without moving left/right will prevent us from getting to the next generation.
Note that in such cases, the score may still slightly increase (maybe a bug in the game that should be fixed)

• Adjust number of input neurons, depending on input
• Play around with number of neurons in hidden layer
• 3 output neurons, one for each class

• gainings
• different fitness function discussion
• screen recording for each fitness function and explain
• describe mutation, selection
• future work: bonuses and stuff

send slides for comments

Let's write down our findings (especially concerning the fitness function) at some point for the final report.

Lets think about a cool fitness function + don't penalize manually :

height/time -> not so good

height²/time ?

norm(height) - norm(time) ?

Goncalo suggested to start with the same network multiple times (100x or so) and observe how networks develop.

• reduce number input platforms
• sort input platforms by height depending on player's position

Let's try, if we can speed up learning by using nodejs or something like that. Also, pay attention to resource usage (updating the dom, storing data, etc).

Player appears on the other side of the game, if it hits left/right bound -> Disable that

• make it clearly visible which network plays which game
• fix css -> breaks when zooming

Right now, when attempting to jump on a platform, the input parameter positions representing the platform are changing multiple times mid-jump or even disappearing completely. That should only happen after a jump was completed.

use less games for training (like 4 to 8)

Would be quite nice to export models.

• dynamic mutation: if stuck -> higher
• dynamic drop out for best players depending on performance

Goncalo suggested to increase the mutation rate. I think we should play around with that.

Display top players on top of the html page.

Disable rendering in order to improve learning speed

Even though Goncalo recommended not to use the best overall player, let's try this out.