Collection of helpfull Maxims

[Hairy Phil]

Don't score if you are point leader

How do you know whether you are point leader on the Random arena?

[pier4r]

I assume with tags and counting. But would be interesting to know.

Also, DrPhoenix, waiting for you in master/grandmaster. We need opponents!

[DrPhoenix]

Well, yes. With Tags and counting. It is not 100% acurate. I think I will have a problem, if one team scores twice on the same click and I can only count up to 5 for each team. Also, you have to do it with team tags to avoid loosing your score if the counting Bot dies. I won't spoil any specifics, but you might guess how I do it, if you play against me and watch the Replay carefully.

[DrPhoenix]

Hope to be in Master or GM soon, if I have time to play. It was easy up to diamond, but now I realize that my sniper AI sucks. Also aiming could be a tat better...

[pier4r]

YOu have plenty of (broken) ghosts to go through!

[sollniss]

It's very easy to count the resources colleced even if they are collected on the same tick. It only takes about 30 nodes to implement.

[pier4r]

From a discussion in telegram. How much stuff is there but doesn't get formalized here or in the wiki.

Zaxon, [04.04.2018, 13:10:20]:
I've now be demoted down into diamond. I hope all these radical changes are worth it in the long run.

Pier A Admin, [04.04.2018, 15:02:42]:
I suggest that you take it slowly . Also I remember I did a quick test after a crash of gladiabots and with 25 nodes without much thinking I was able to hold around 1450. (Even the ghost is still holding in that range )

So if you may go quite down a possibility is to rename your ai as a backup and start new. Maybe with the experience you have you avoid bringing with you the technical debt.

Sometimes I think I should do the same but I like the cleaning up challenge

Zaxon, [04.04.2018, 16:05:46]:
Yup, I've got my previous AI saved. How can you tell the levels of previous ghosts? Just browsing the Latest Matches screen?

Zaxon, [04.04.2018, 16:06:30]:
Which would be quicker to do?

Pier A Admin, [04.04.2018, 16:07:48]:
as far as I checked yes, moreover “more” it is not that much needed.

Pier A Admin, [04.04.2018, 16:09:40]:
It depends. If you messed up badly and you know you changed 50+ nodes, starting again.

If you messed up not so much nodes and you kept a log of the changes, fixing the AI that you have.

If the AI in general is around 100 nodes or less and your score is under 1700 and 300-400 points away from your maximum. Starting again.

So if you have a not so big AI, and you are 1400, I would start again fresh. Also make a plan. I mean gladibots is awesome for so many organizational concepts it teaches.


Make a plan.
Fix the main goal.
Set short term goals.
Take it slowly.
keep a log of conceptual changes so you know what you touched/did.
Keep a log of observed patterns, like “oh, my bots should avoid this case. Or should do this”.

You can use paper, or apps like one note.

[TomCat]

- Don't score if you are point leader

And how do you know you're point leader? I am just curious.

[ElvenMonk]

Well, yes. With Tags and counting. It is not 100% acurate. I think I will have a problem, if one team scores twice on the same click and I can only count up to 5 for each team. Also, you have to do it with team tags to avoid loosing your score if the counting Bot dies. I won't spoil any specifics, but you might guess how I do it, if you play against me and watch the Replay carefully.

Well actually it don't have to be team tags anymore, and all bots can count by themselves. And you can count score up to 32 for each team, and can track each unit scoring on same tick independently.

So my maxims are:
- measure every possible distance and tag every object accordingly at the first tick to define strategy
- measure global time and use it when game just started or about to finish
- count when bullets will be fired to avoid damage when possible and so that you don't start fleeng on last aiming tick
- count the score to go mad when you need to
- team tag everithing nearby to know if teammates could help
- locate at least zones of the map to always know where to go and where to flee

[pier4r]

That's an extensive use of tags! I imagine the time needed for it!

I'm still tagless and it takes a lot of time nonetheless.

[masterplayer]

So my maxims are:
- measure every possible distance and tag every object accordingly at the first tick to define strategy

IMAO that's not necessary. Focus on one strategy.

- measure global time and use it when game just started or about to finish

How do you count global time?

- count when bullets will be fired to avoid damage when possible and so that you don't start fleeng on last aiming tick

IMAO that's not necessary. Improve your retreatment.

- count the score to go mad when you need to

Yes.

- team tag everithing nearby to know if teammates could help
- locate at least zones of the map to always know where to go and where to flee

Why would you do this?

[ElvenMonk]

How do you count global time?

By combining tags on 2 objects. 2^10 ticks ~ 4 min 16 sec.

[pier4r]

Repeating sometimes helps.

- spam games to get things to change, the elo goes the elo comes back. Until one doesn't dominate, there is no problem in spamming. Of course if one doesn't break his ai with changes.
- Focus on avoid losses that takes 13 points of more from your score. All the other losses are allowed. (rationale: you may grow and losses that were "acceptable" become not that acceptable anymore)
- Once one doesn't have any more losses taking away 13 points from him, focus on keeping small the percentage of defeats (if one plays 20+ games in a session). That is, if one plays 54 games with 14 defeats, the percentage is almost 26%. Try to convert a defeat to a draw at least and then repeat the process. Of course victories are even better but draws are a good start.
- don't change too much at once, unless you have everything under control. A small change every 5-10 games is quite good to evaluate if it is neutral or improving things rather than breaking things down.

[Chilio]

My little addition:
Know when and how to attack.

Start attacking to either push or to stop walking to get into enemy territory.
Switch attacking only when a higher priority target comes up: one that has enough priority to stop aiming on the current target.
Most of the time, instead of switching targets, stop attacking and move for one tick to start attacking the next tick again. (For example when your current target is dead or out of range, but just when you're not too far into enemy territory.)

[pier4r]

Before the unranked league was reality, some vocal players (ntk) said that people would play only unranked and almost never ranked.

Well in unranked one can find enough good enough ghosts to improve is AI or test it quite a bit but the difference in difficulty between ranked and unranked is quite impressive.
In general in unranked one finds ghosts that are always less strong than their ranked counterpart.

So my suggestion is:
- if you don't mind losing points as you will win them back, test ranked.
- do you want to test a radical change, to catch at least the most obvious bugs? Unranked is ok.
- do you want to refresh the status of your ai after a period of inactivity? Unranked is ok.
- do you want more variety? Unranked is ok.


Today I devised the following approach:
1. play ranked.
2. get a point where you need a fix.
3. if the fix is minor, go to 1
4. if the fix is not minor, go unranked first, play 10+ games to have a feeling of the soundness of the fix.
5. if the fix is not sound, go to point 2
6. if the fix seems sound, go to point 1.

In this way one plays quite relaxes, without the little negativity of score changes.

[pier4r]

So my suggestion is:
- if you don't mind losing points as you will win them back, test ranked.

I want to expand on this because I formalized the approach a bit better. I know that losing points is a blow for the ego, so if you cannot go through it (at times it happens to me too), the following is not for you.

If instead you are in a mental state where you don't care about points, ranked is where - so far at least - there are the most optimized ghosts because those not optimized ones are quickly disabled. Plus the matchmaker is great. Unless you are at the top (or top bottom) the matchmaker will present you countless (30+) of opponents with a variety of approaches.

So one has to establish first the major goal. It could be: win as much as you can, or lose as little as you can (so draws are ok). I use the second latter.

One can also observe that the matchmaker, at least at the moment, other than providing a great variety of opponents, ensures that we get opponents "close" to our current strength. So optimal for testing, as getting opponents too weak or too strong often lead to no useful data to improve one's AI. Either the opponent is not that difficult, or it has so many subtle actions (because it is too strong) that it is difficult to understand where one fails.

Thus this suggestion is based on this two strong points for the matchmaker: variety, and variety "that we may understand to improve the AI".

Then each played match can fall in one the following categories (well actually a match can be of category b and c at the same time):
-(a) it is a good match. Likely your AI has problems but the course of the match was satisfying for you and you won't check for flaws.
-(b) It is a match that provide actionable data, you are quite sure where is the problematic part in your AI that need to be fixed. (you may be wrong, but the feeling is that you are quite sure).
-(c) It is a match where is not that clear what went wrong. You need to collect more matches like those to then decide, but you feel you can crack the problem.
-(d) It is a match were something went wrong but you have no clue were, unless you go through the match multiple times in normal and slow motion.

The suggestion is to ignore (unless you have plenty of time) the type (d) matches. The type (a) are ignored by definition.

Therefore using ranked matches as testing, ignoring the score, helps to get actionable data or data that will be processed to extract clues or general patterns. Relying on the score to be sure to get "useful opponents", and nothing else, will likely produce a lot of match of type (b) and (c). The data from those, when used carefully, can lead to pretty neat improvements of the AI. If nothing else, at least one has data to lower the amount of small problems that may help in the long run.

[pier4r]

Another suggestion that may be forgotten even by seasoned players or problem solvers.

If you play frequently and change every now and then your AI, you may do little changes (affecting few nodes). From time to time you may introduce something that affects (or consist) of more nodes, say 10-20 . Anyway always playing do not give the possibility to invest time in larger changes.

The point being, another way to see multiplayer games is:
1. multiplayer games: to test small changes (few nodes)
2. take a pause - that is: work without often playing online - to introduce a larger change (10+ nodes) and then refer to 1.
3. take a longer pause for more massive changes (20/30+ nodes). Then refer to 1 or 2.

I observed that when I have time and I can play, I end up spamming games but practically changing little parts. While sometimes improvements (or refactoring) requires more time working in private to then test online.

Though little changes may still produce improvements from time to time.

[pier4r]

I'd like to add that a good point is to keep your code modular and easy to clean up if one accumulates too many structures (and technical debt).

The game changes at every major version and even if it wouldn't there is so much to explore, see viewtopic.php?f=7&t=1749 , that most likely after few weeks or months the strategies that were winning aren't winning anymore (especially with the availability of ghosts that allow for faster counter development than in the past),

Many, like me, build a huge decision tree (what we call also AI) that then it is difficult to change in something more fitting if we realize we have other ideas. To start from scratch is overwhelming. Some players did it successfully many times. Mumpsimus (going up and down the scores), ritter (with new accounts), gtresd, gilberreke and others.

Anyway for some may be too much (I feel that way in this moment). Especially after long periods of inactivity it seems one need a lot of time to be competitive again. It may not be always the case that the AI is outdated, for example I stopped playing in Nov 2017 and I restarted in Feb 2018 and my AI was still on the top, but it can also be the case.

Therefore to avoid having this inertia in a direction that you don't like anymore, organize your code to be able to swap completely new elements in if it is possible.

My code is partially modular (I use subtrees but modularity is not just that, as subtrees may have a connection with each other), and therefore to start from scratch again I estimate a bit of time investment again (some 20-30 hours) and that keeps me from playing a bit. Because it is time I don't have in the moment.

Hence this advice. Try to keep yourself modular and keep your AI manageable. Imagine yourself coming after some months of inactivity and just quickly developing fixes without the need to scrap your code that, although using subtrees, may feel monolithic (my current situation=.