It appears like most of the top tactics (in OTL etc) are variations of the same basic design. How boring.
To change this, I thought I'd make short guide on how to build a tactic from scratch and good enough to hit top 20 in OTL.
This is a guide without any tactical advice. Just basic scientific methodology applied to CM :-)
You need: a fast PC, a quick holiday benchmark save, and a lot of patience. Also, some automated way to gather results from holiday benchmarks is a huge time saver. Also a tool (like cmtactool) for assessing benchmark resulst is needed.
What I did this weekend (actually, most work was done my computer, I only poked at it every now and then) was to test how hard it was to beat top 20 in OTL. And it was not very hard.
A notch is the smallest unit you can move a player on the wib-wob screen.
The pitch is the whole playing field (duh). The pitch is divided into 12 squares where positions can be given with and without ball. A square is 9x9 notches, and the whole pitch is 27x36 notches.
Step 1. Overview
Build basic formation on the overview screen. Place out the players on the overview. And make arrows where you really want them.
Step 2. Basic wib-wob
Basic wib-wob. This is to create a starting point, it does not have to be perfect, or even have significant amount of effort put into it.
I never liked the default positioning of the players, so I kinda move them to my own default locations before I start.
The wob screens are usually okayish, only some minor adjustments here (like moving the left hand squares, I leave the right wingers where they are, but move the two central players 1 and 2notches to the right, and the left wingers three notches to the right).
The wib screens are based on the wob screens, defenders are pushed around 3 notches up, dms 3-6 notches, midfielders 6-9 notches, and attachers 6-12 notches from their positions on wob screens. Also I spread out fcs and amcs a little on wib screens.
The result I save with a short name like: "451".
Here the tactic building starts.
Step 3. Player instructions
The first I do is to find a basic, decent, setup of player instructions.
I build up to ~8 versions of the tactic, with different forward runs, run with ball, and set pieces instructions. Naming them like 451_a, 451_b, ..., 451_h or smth.
Benchmark each some 50-200 times, until you feel you have a candidate setup that you think is best.
For example, in my experiment this weekend, the chosen setup was the first setting, i.e the 451_b tactic.
Step 4. Player positions
This is the main step.
Start with your candidate best version of your tactic (initially the one from step 3).
Do only one change, and save it with a new name. Like 451_b1. Then you go back to your candidate, and do another one change. Save it like 451_b2 or smth.
So what is one change? Depends on. It usually is a group of players (say MCs) that are moved a few notches up or down on a wib/wob square. The smaller a change, the more benchmark runs you need to run. The larger the change, the greater the chance that you don't distinguish a pattern (some part of the change is good, another bad).
A rule of thumb is to start with larger changes (totalling of 20 notches or so), and then gradually go down to changes of 3-6 notches.
A change of 6-12 notches takes about 200-500 benchmark runs to evaluate, and small changes like 2-3 notches can take 3000 (and sometimes that is not enough).
Since my CPU has 4 cores (8 virtual) I can run up to 8-9 CMs at the same time. So I make up to 8-9 variants of the candidate tactic, naming them with another new letter / number appended to the name. (Usually, doing 2-3 changes is enough to find an improvement). To use a computer science term, I'll call these changed tactics children of the candidate tactic. Run these until you feel you have found one or more changes that improve on your candidate (or determined they are all just worse).
In case of one improved version:
- Then the improved version becomes your new candidate tactic.
In case of all worse:
- make 8-9 new changes.
In case of several improvements:
- cut and paste all improving changes into a new "consolidated" version. Benchmark the consolidated version as well, to verify it is better than the previous candidate and its children. (If not, cut and paste together any improved version, omitting some of the changes).
- sometimes the changes that improved cannot be applied at the same time. For instance, you might have moved MCs up in one version, and down in another -- and both are better than your candidate. Then pick one of them, and maybe return to the other variant later.
(In building Iodine, this happened 4-5 times, and all the changes gave birth to interesting tactics in the end).
Step 5. Repeat step 4
Do step four until you give up. This weekend I had to do step 4 around 9 times to beat tactics on top 20 OTL, and after 11-13 times my tactic was as good as #10 OTL.
Occasionally you can go back to make a change in player instructions, and see if it changes anything.
Patience. Benchmark more.
- Use CM tactool to determine when a tactic really is better than another.
- Tight cases (<0.10 pts diff or so) also look at conceded and scored. A tactic that scores slightly more points despite scoring less and condeding more is likely a the worse one in tight calls.
- Make your new candidate tactic prove they are worthy. I often run the new leader 2000 seasons or so to make them prove they are the best so far. Picking the wrong candidate tactic can cost you a lot of time (you'll notice it as nothing improving on it ... because your impression of its performance was luck and not its actually scores).
- If you can parallellize your CM runs, you can do 2500 seasons of 8-9 tactics while you sleep. Use that time to verify your findings.
- An improvement of over 0.5 points (or more) in average is very good. You should expect improvements of around 0.2-0.3 points per change.
There is a lot more to say, but this is just a forum post, not a book.
Last edited by Tapani; 29-02-16 at 11:51 AM.
where do i find cmtactool?
i go there but i couldn't find it
Tapani. Wow. This is amazing and exactly what I have wanted to read for ages. I've been trying to create a new tactic all week without much luck so this will definitely help. I have my own method of collating data using a spreadsheet I designed but how do you run 9 CMs at the same time? I've been using your benchmarking save which is brilliant for benchmarking but how do you get more than one instance of CM to benchmark at the same time?
Tapani, It might be a question about pretty cosmetic thing, but, from your experience, can you feel the difference in obtained results between the DB size when the tests are conducted?
Nice one! After using other members' tactics and winning everything there is to win a lot of times I've been wanting to make my own for some while. Never seem to find the time nor the effort to do that. May give it a go in the future though.
You tactics makers: how long time does normally pass from the point where you first start on a new one until you are satisfied with it (then you can always make more changes ofc)?
You need to use the dev mode feature. Also you need scripts that adds a delay between the instances, parses the results etc.Originally Posted by Jesus
It is easier to make several instances running separate tactics, but that also needs a setup of several CM folders.
For me ... I often tire of a tactic after a few days. To get top results (Iodine) ... maybe might have taken a month (single core CPU time, not my time).Originally Posted by felxsjo
No such thing as perfect tactic will be made. Too many variables to test, so I was never satisifed. Ever. .
This is good read, but... I thought most of that was known by now? The Dugout started this after all, though at that point in time, no-one really knew what they were doing with tactics. It was trial and error, but when DDL came into play, I think, some people made breakthrough back then already.
Not that you see DDL concept used nowadays. Might actually go back to it. Just rambling .
Thank you so much for this Tapani! The timing is great because I have been trying to create some interesting tactics and trying to make them work. It looks like I will have to finally figure out how to work this dev/benchmaring mode. Does anyone have the De Graafschap benchmark template save game?
So I've started using this method tonight and running CM0102 in parallel is my biggest stumbling block. I don't know how to get CM0102 to load up the tactics individually for each save. As a result I'm running 50 tests of each tactic independantly which feels like a longer process than it should be..
So I've tried multiple work around tonight to try and run CM0102 in parallel but haven't got anywhere.
First I copied new folders of CM0102 to my desktop. I then renamed the folder, the benchmark .exe and save files. This didn't work.
Second I installed fresh copies of CM0102 on to my hard drive, giving them unique names. I then renamed the benchmark .exe and save files. This didn't work.
Third I created new partitions on my hard drive, installed fresh copies of CM0102, renamed the benchmark .exe and save files. This didn't work.
The problem I have found is that I get this error:
If not PM Cam or Faz for it.
Oh look another person I helped... funny that as two very senior members of this forum just ignored this post
Sorry guys for removing your bickering - have your barfight in the pub.