Talk:Movement

From Robowiki
Jump to navigation Jump to search

Contents

Thread titleRepliesLast modified
Doing better against simple targeters1603:30, 7 October 2017
Gradient Descent210:12, 8 August 2017

Doing better against simple targeters

I'm trying to make my bot better against Wave Surfing Challenge bots and I tuned the movement for a very long time but I can't get more than %90 C against C and %93 against B. Does everybody have this problem when making a new bot or It's me the problem?

Dsekercioglu (talk)10:39, 6 October 2017

My avg against C is like 85% or so. I have never extensively tuned against 2K6, so I didn't expected it to be good, but I always had bad results against GrubbmGrb VG in 2K7, so I suspected it would do bad.

At the same time it shows it is perfectly possible to get a reasonable movement even without doing well against those bots (you can certainly hit top 30 at least with that), but I would recommend not doing so. Now I'm trying hard to do better against those bots without losing all the points my movement earned me and the experience is justing frustrating.

Rsalesc (talk)11:02, 6 October 2017

AFAIK, you can crush LT/CT guns by moving closer that there’s only one wave in the air. Then some form of fire situation design will also give you some chance to generate Stop&Go movement and the alike, that works even against top guns.

Xor (talk)11:41, 6 October 2017

Can you run your bot against those three when you have time? Just for the sake of my curiosity :P

Rsalesc (talk)11:43, 6 October 2017

ScalarBot is doing very bad against wsc LT/CT, but my score against those bots in real world seems to be even higher than some top bots. Seems that my gun is destroying them before they even had a chance to shot. ;)

I tried some dedicated buffer for the simple targeters, but it just don’t work (not giving me more score).

I think that’s because I’m too far away, visiting every fire situation, and making simple guns fire at every possible gf.

Xor (talk)11:51, 6 October 2017
That is probably because bots using "Fast Targeting" also use some non-learning movement.
Dsekercioglu (talk)12:29, 6 October 2017
 

But since simple guns are pretry predictable, you can always crush them by preloading gfs.

Xor (talk)11:52, 6 October 2017

Yes, I do that, but I eventually start getting hit and then I stop falling back to my preloaded GFs. What I wonder is how those top 2K6 bots can still learn well.

Rsalesc (talk)11:55, 6 October 2017
 

Just had the time to run that ;)

| ScalarBot 0.012l40.mc || Author || Type || 100.0 || 88.82 || 89.56 || 92.79 || 100.0 seasons

I haven't run it for a long time, and the last time I run it, it seems to score ~75 against WscB. Anyway, it's way better than I thought, but still way behind Dookious.

Xor (talk)02:51, 7 October 2017

Wow, it seem that SimpleBot is performing way better —

| SimpleBot 0.024c.c.mc || Author || Type || 99.42 || 95.18 || 94.67 || 96.42 || 100.0 seasons

Where the biggest difference is the lack of precise intersection, and the use of 1 / (1 + abs(x)) as danger function.

Also, SimpleBot is examing stationary danger in both waves.

Xor (talk)03:13, 7 October 2017
 
 
 
 
 

Gradient Descent

Can't we use gradient descent for movement in melee? Every bot would have a function(Gaussian), bullets would have a 1d gaussian function. Then we would get the 2D derivative and move that way.

Dsekercioglu (talk)08:02, 8 August 2017

It's an interesting idea. I am not sure if you mean it this way, but I think it would work almost the same way as Anti-Gravity_Movement?

Cb (talk)09:35, 8 August 2017

Yes, they are very similar. One of the only differences is momentum.

Dsekercioglu (talk)10:12, 8 August 2017