Reinforced Learning 4: Bugged or Impossible

[Srint]

Got the game 2 days ago and loved it, finished with gold medals on everything on two days of just playing (seriously LOVED it) but I am stuck on RL 4.

The front lidar is just two lines long, not big enough to avoid any colisions whatsoever when driving above 53 km/h on the minigame (even though the lidar on RL 0 is like a mile long :P )

I have no idea what the extra classifications are used for (object to car / road block) and after 6 hours of testing (seriously) I found that they do absolutely nothing. Might be wrong, but just by placing a node on the schema, without connecting it to the car's logic board, the car can magically drive with no input :P.

Finaly, the sliders are total RNG, and I still don't get why they are used for. Sometimes, increasing something will help. Othertimes, using a 10-100-1-10-10-x-x system will provide the best results in training, but I still can't get past 58.41 km/h.

How am I supposed to reach from 58.41 to 62??
Has anyone done it? Is there any way to increase the front lidar's detection range?
Even a better description than the game provides on the RL systems would seriously help (sliders etc.)

[Legolas]

I'm having the same issue, stuck at 57 something at best in this task... Too bad because I got gold everywere else, except Reinforced Learning 2: I got bronze first, then improved it to gold but the task tree just won't update its status to gold, unlike regular tasks. I hope we will find a solution, this game is awesome :)

[Switch]

I'm feeling like there should be a consistent tutorial about Reinforced Learning on the game. It's the only part where I don't know exactly what the knobs does or doesn't. Seems like a random part of the game... for a logic game, it's strange

[GospodinNoob]

Originally posted by Switch:

I'm feeling like there should be a consistent tutorial about Reinforced Learning on the game. It's the only part where I don't know exactly what the knobs does or doesn't. Seems like a random part of the game... for a logic game, it's strange

Hi!

This part of the game is a little random, that's true, just because we add a real machine learning inside the game. But you can stabilize your model by tuning the model parameters.

[GospodinNoob]

Originally posted by Legolas:

I'm having the same issue, stuck at 57 something at best in this task... Too bad because I got gold everywere else, except Reinforced Learning 2: I got bronze first, then improved it to gold but the task tree just won't update its status to gold, unlike regular tasks. I hope we will find a solution, this game is awesome :)

Thanks for the bug report about medals. will be fixed

[GospodinNoob]

Originally posted by Srint:

Got the game 2 days ago and loved it, finished with gold medals on everything on two days of just playing (seriously LOVED it) but I am stuck on RL 4.

The front lidar is just two lines long, not big enough to avoid any colisions whatsoever when driving above 53 km/h on the minigame (even though the lidar on RL 0 is like a mile long :P )

I have no idea what the extra classifications are used for (object to car / road block) and after 6 hours of testing (seriously) I found that they do absolutely nothing. Might be wrong, but just by placing a node on the schema, without connecting it to the car's logic board, the car can magically drive with no input :P.

Finaly, the sliders are total RNG, and I still don't get why they are used for. Sometimes, increasing something will help. Othertimes, using a 10-100-1-10-10-x-x system will provide the best results in training, but I still can't get past 58.41 km/h.

How am I supposed to reach from 58.41 to 62??
Has anyone done it? Is there any way to increase the front lidar's detection range?
Even a better description than the game provides on the RL systems would seriously help (sliders etc.)

Originally posted by Legolas:

I'm having the same issue, stuck at 57 something at best in this task... Too bad because I got gold everywere else, except Reinforced Learning 2: I got bronze first, then improved it to gold but the task tree just won't update its status to gold, unlike regular tasks. I hope we will find a solution, this game is awesome :)

Some hints: You can reduce the lidar area, by clicking on the lidar cells in the lidar window. Try to create more "sparse" observations. It will increase the learning ability of the model.

[Switch]

Originally posted by GamineBrownies:

Originally posted by Switch:

I'm feeling like there should be a consistent tutorial about Reinforced Learning on the game. It's the only part where I don't know exactly what the knobs does or doesn't. Seems like a random part of the game... for a logic game, it's strange

Hi!

This part of the game is a little random, that's true, just because we add a real machine learning inside the game. But you can stabilize your model by tuning the model parameters.

Thanks for the response, I understand that needs to be random, what I fail to see is how tuning the parameters affect the output, the tutorial is falling short for me in that aspect, and getting my hands on testing doesn't quite reflect that either. I passed those levels just clicking random params and crossing my fingers. For example, at one time the car started slowing down until it stopped and I just fed it an acceleration sample, with dodging included, so I was left astonished as to why and how to *correct* that.
In the final part of the game (last 20%) I just end up doing random stuff rather than knowing what I'm actually doing.
It's a good game for sure for someone who likes programming like me, but made easy so it's not working tedious, also I got to learn quite a lot about neural nets. Keep up the good work!

[lukaslew]

I would be nice to know if settings should be changed for each training session ;)

[Imi]

Sadly, the easiest way to "pass" this is to switch every lidar off, then don't bother with Training and keep resetting and Try-Run immediately until you got lucky.

[your.sheepy]

One thing I suspected is that perhaps most or all lidar signals are merged into a few streams, if not one stream, in which case the side lidar signals would "drown out" the front lidar signals, so the car can't learn when it should accelerate.

What I did is turning off the side lidars, leaving only the top row, and I consistently get a profitable model. Oh and I always increase population to max.