My feelings about gladiabots tactics since alpha 4.X

pier4r
Skynet
Skynet
Posts: 3390

My feelings about gladiabots tactics since alpha 4.X

Post#1 » 25 Dec 2016, 19:26

The feeling was emphasized in alpha 5.2 but I still believe that whatever fight between AIs under 100 nodes (per bot) may show some AIs stronger than others, but not all the others (well not so much experience in alpha 6.1.3 but from the reports I think it is still valid).

Like this:
[youtube]zzKGnuvX6IQ[/youtube]
http://www.reddit.com/r/Gladiabots/wiki/players/pier4r_nvidia_shield_k1 -> Gladiabots CHAT, stats, insights and more ;

User avatar
HBomb
Algorithm
Algorithm
Posts: 93

Re: My feelings about gladiabots tactics since alpha 4.X

Post#2 » 25 Dec 2016, 21:20

I had a hunch along these lines, but thank you for formalizing and wording it properly. I guess this is also a mathematical reason why the newcomers tend to win and climb, and oldies must completely change to keep up.

pier4r
Skynet
Skynet
Posts: 3390

Re: My feelings about gladiabots tactics since alpha 4.X

Post#3 » 31 Dec 2016, 15:34

Another feeling that I often have is: "ok, i reached the best that I can, if i want to improve I should put a lot of effort for a little refinement not so noticeable", the day after: "Oh my god, why did I write those poor conditions in my tree?". I change them and I get a quite noticeable improvement with relatively little effort.

And this happens, like, always. So far I was never able to reach a point that I can say "to get better, and not a little, I should create way more refined optimizations".

Edit: another one is - "how many options that I can choose but they are not so useful", the next day I thank GFX for those options that suddenly I need desperately.
http://www.reddit.com/r/Gladiabots/wiki/players/pier4r_nvidia_shield_k1 -> Gladiabots CHAT, stats, insights and more ;

Hamster
Script
Script
Posts: 28

Re: My feelings about gladiabots tactics since alpha 4.X

Post#4 » 08 Jan 2017, 18:53

I think right now there is an idea (which seems intuitive) that the way to build good AI's is to get something basic done and then add more and more nodes that cover all the niche cases. We work from high-to-low level.

However, its always possible that some other player comes along and decides to something different at the high level; if it works, there are massive gains in performance for that player (compared with someone tweaking low-level stuff).

For instance, its very easy to decide if a bot losses its shields it should retreat. We then add 100's of low level nodes that decide how that bot should retreat. The problem is that as we build all these 'corner case' nodes we never end up questioning the "high-level" premise; "should the bot retreat in first place?"

maybe you decide it should retreat some of the time, so you write 100's more nodes covering the niche cases where you don't want to retreat. But this once again falls into the 'low-level' optimisation trap; we fail to question whether the general principle of "retreat some of the time" is actually correct.


Maybe this is why the game can feel a bit like magic dice at times? The new players that sail through the ranks are probably not doing huge amounts of low-level stuff. I say that because functioning AI's 100's of node deep take a long time to build and in several cases require detailed knowledge of bot stats and game mechanics.

If the new players are not diving too deep, but are, on occasion, beating/drawing top players then the obvious conclusion is that these players are innovating at a high(er)-level instead.

==========================================

As an interesting experiment, can you guys build an AI with, say, less than X no. of nodes that can consistently beat your main AI's?

If you can do that but then find you cannot improve your main AI (without worsening performance in other match-ups) that would suggest the game is fundamentally rock-paper-scissors. If you can beat this simple AI without harming other match-ups then great! you have improved your main AI!

If you cannot build a simple AI to beat the main AI then that either suggests you have yet to discover a "high-level" strategy that works or that no such strategy actually exists. In either case, it seems right to try and focus attention on low-level stuff.

pier4r
Skynet
Skynet
Posts: 3390

Re: My feelings about gladiabots tactics since alpha 4.X

Post#5 » 08 Jan 2017, 20:44

Nice point of view. I agree the part at higher level and improving Ais with less than X nodes
http://www.reddit.com/r/Gladiabots/wiki/players/pier4r_nvidia_shield_k1 -> Gladiabots CHAT, stats, insights and more ;

User avatar
Ritter Runkel
Neural Network
Neural Network
Posts: 498

Re: My feelings about gladiabots tactics since alpha 4.X

Post#6 » 09 Jan 2017, 06:06

Very good idea. I like this approach to decide wether to optimize or to restart.

pier4r
Skynet
Skynet
Posts: 3390

Re: My feelings about gladiabots tactics since alpha 4.X

Post#7 » 09 Jan 2017, 22:22

Sometimes I have a streak of many victories that bring me +4 or less points (nullpointer knows this well I suppose), and then I think "stop with this, bring the heavy weight, I want blood!". fortunately with the new matchmaking I do not have to wait much for it.
http://www.reddit.com/r/Gladiabots/wiki/players/pier4r_nvidia_shield_k1 -> Gladiabots CHAT, stats, insights and more ;

pier4r
Skynet
Skynet
Posts: 3390

Re: Avoiding movement: umbilicated cialis instructions manoeuvre: generic celebrex stresses surgeon turmoil.

Post#8 » 17 Jan 2017, 20:09

ifuvqidecna wrote:spam


Profile:
Are you a spam bot?:
Nope, I'm a nice human :)

You liar :)
http://www.reddit.com/r/Gladiabots/wiki/players/pier4r_nvidia_shield_k1 -> Gladiabots CHAT, stats, insights and more ;

pier4r
Skynet
Skynet
Posts: 3390

Re: My feelings about gladiabots tactics since alpha 4.X

Post#9 » 21 Jan 2017, 15:04

Anyway another thing that is common in gladiabots is that it teaches well the https://en.wikipedia.org/wiki/Test-driven_development . Unless one tries to sort out all the problems in the test arena, the ones that play online and find fixes are following the test driven development path.
http://www.reddit.com/r/Gladiabots/wiki/players/pier4r_nvidia_shield_k1 -> Gladiabots CHAT, stats, insights and more ;

User avatar
NullPointer
Autonomous Entity
Autonomous Entity
Posts: 539

Re: My feelings about gladiabots tactics since alpha 4.X

Post#10 » 21 Jan 2017, 15:28

I believe TDD would be something like this:

- come up with an idea
- build a test AI that your idea is supposed to be strong against
- build an AI with your idea
- put them to fight
- depending on the result, refine the test AI or the main AI
- put them to fight
- repeat last 2 steps until you're satisfied
- then you put it in production

If you go directly to the last step, you're not following any testing strategy at all.

pier4r
Skynet
Skynet
Posts: 3390

Re: My feelings about gladiabots tactics since alpha 4.X

Post#11 » 21 Jan 2017, 15:52

NullPointer wrote:I believe TDD would be something like this:

- come up with an idea
- build a test AI that your idea is supposed to be strong against
- build an AI with your idea
- put them to fight
- depending on the result, refine the test AI or the main AI
- put them to fight
- repeat last 2 steps until you're satisfied
- then you put it in production

If you go directly to the last step, you're not following any testing strategy at all.


First: nice summary.

Second: it depends what is your "build a test AI". If my test AI is "idle" (or often my old AI) and I'm satisfied to beat it, then I use the multiplayer as testing.

Production is something that should be stable for a while, while in multiplayer mostly we have continuous changes. For me multiplayer is not a production. A tournament with AIs that cannot be changed would be production. Or a tournament with stages, before the next stage the AIs are fixed.

But yes if one does only the last step, assuming the multiplayer as production you are right.
http://www.reddit.com/r/Gladiabots/wiki/players/pier4r_nvidia_shield_k1 -> Gladiabots CHAT, stats, insights and more ;

User avatar
NullPointer
Autonomous Entity
Autonomous Entity
Posts: 539

Re: My feelings about gladiabots tactics since alpha 4.X

Post#12 » 21 Jan 2017, 21:45

I think that regardless of what "production" means, a core concept of TDD is to write the tests before the production code. Given that, even if multiplayer is not "production", it doesn't mean you're following TDD if you're not actually writing any tests for your AIs before you create them.

By the way, congrats on your high scores, you keep floating above 1800+, what a boost from the previous Alpha versions.

pier4r
Skynet
Skynet
Posts: 3390

Re: My feelings about gladiabots tactics since alpha 4.X

Post#13 » 21 Jan 2017, 22:07

I thought about this a bit more and I see it in the following way. Imagine I cannot produce all the test but I know that there is a lot to test, then I use beta testers (however I define them). That is still not production for me, but those are "unknown tests" that I will clarify with logs.

So if I see multiplayer as "now I do my definitive AI", then it is not really testing. If I see it like "ok let's improve the AI until I have the definitive AI", then it is like I send my AI to beta testers.

If one instead use the strict definition "writing a test first" on your own, then yes, there is no TDD.
http://www.reddit.com/r/Gladiabots/wiki/players/pier4r_nvidia_shield_k1 -> Gladiabots CHAT, stats, insights and more ;

Return to “General Discussions”

Who is online

Users browsing this forum: No registered users and 1 guest