Dominik Nitsche recently posted a https://twitter.com/DominikNitsche/s...639770112?s=20on Twitter calling out Odin for it's lack of accuracy in it's solves. Apparently, it only uses 8%. This sparked a few questions for me:

1.) If two solves were set in front of you and the only difference between them was the accuracy, could you tell which was which?

2.) If a human player could execute a strategy provided by a solve using 8% accuracy, how strong of a player would they be? Could a human player beat this strategy?

3.) If the solve using 8% accuracy is weak(er), could you identify the leaks and exploit them? (Let's say the entire game of poker were boiled down to this one solve (one board/same parameters), it should seem much easier to identify and counter mistakes. You'll have complete knowledge of your opponent's strategy and they won't adjust.)

4.) For 3, what would your process be for identifying these mistakes?

1) A lot of the time yes, converged sims tend to be pretty clear cut while non converged sims mix a bunch of unnecessary/bad ****.

2) A human can do better than 8% EV loss as is. I guess you could be a winning player if you copied this sim and played a soft pool but it would be very silly as you could just have higher EV by playing like a good human player and learning those sims would take longer than getting good.

3) Probably not possible for a human to wing exploits against the bot, but at least for a lot of the flop strategies that the bot misplays like for instance donking flops that shouldn't be donked, you could run nodelocks off table and find how to make a lot of money there. Probably possible for a human to choose blockers very marginally better and avoid taking lines that shouldn't be taken.

4) Mostly running nodelocks of common mistakes the bot makes, assuming as you said that you know his entire strategy.

BTW I don't think the sims are run to 8% accuracy, I don't believe anyone can be that dumb. I'd be surprised.

I wouldn't be surprised. The lower accuracy could allow them to get to market faster and give them some room for future improvements to advertise. Also, I doubt the average consumer will even notice or care that much.

The sims would just be donking all over the place, probably every single spot. It'd be way too obnoxious imo

GTO+ solved to 8%:


One of the tell-tale signs of low accuracy is that the strategy/equity distribution looks clumpy, and strategies are overly mixed:


Odin sim:


GTO+ solved to 0.1%:



I don't know if they solved to 8%, but it's pretty clear that they didn't solve down to high accuracy.

The advantage of solving to higher accuracy is that strategies become less mixed and easier to implement and understand. It's not important if a human could or couldn't beat low-accuracy sims, it's a question of maximizing your study time.

8.0% is straight up worthless.

1.0% is the start of becoming acceptable

0.5% or better is great

0.1% or better is essentially perfect

Does it matter much if I run my GTO+ sims to 0.5%, 0.25% or 0.1%?
Are there specific spots where I should run them deeper? How do I recognize these?

Short answer is no. The difference between the accuracy of 0.5 and 0.25 and then 0.1 is very low

I think it is a very "it depends" situation. Convergence of these algorithms is not completely understood.
Many variations and tweaks exist to improve convergence.

If you solve to 8% how do you know if you are converged?
Solve to 4% and see if the strategy has changed.
If it is still changing solve to 2%. etc.

Some of it is experience from using the specific tool.

The simpler the scenario you set up the better it will converge (usually).
There are always going to be pathological cases that converge much more poorly.

I think a lot of the current tools are trading off the time to produce an answer (whisper it - real time assistance)
but they are trading that against accuracy.

Different tools will converge to nearly identical solutions IF you let them solve to high enough accuracy. The lower your accuracy, the more correct answers exist. Many strategies can achieve less than 8% exploitability, fewer strategies can achieve less than 0.2% exploitability.

The level of accuracy you need largely comes down to your use case. Some things require higher precision than others.

• 1%+ Literal garbage
• 0.5% Aggregate reports / larger trends
• 0.25% Play against the solution
• 0.1% Optimize bet sizes or compare EV of different strategies

The other thing that comes to mind regarding the accuracy are the parameters used (ranges and sizings). Many inaccuracies can arise before the solver even begins running (poor ranges, limitations on bet sizing options due to computational power, bunching, etc.). I wonder what impact these things have on the overall solve compared using a lower accuracy target. If you could choose to have perfect parameters but only 8% accuracy OR inaccurate parameters with perfect accuracy, would either be a better option, or will both just be effective garbage.

Sure, studying good information will be best. However, if this low accuracy solve is producing a very weak strategy, which some will say is garbage, shouldn't a human player be able to easily identify the flaws and be confident in beating it?

Looking at the low accuracy solve you provided above, even with all of the problems you mentioned, I'm not sure I see how a human player could defeat this solution without using some sort of assistance. I'm not saying it cant' be done - perhaps the solution just looks intimidating.

I wonder how many players would question it? My money would be on players seeing a significant increase in donking happening at the tables.

8% is really bad so it's not a great comparison. But you bring up a good point about optimizing your computing power for the gamespace vs accuracy. I don't think this is well-researched.

Again I think it really comes down to the use case. For example, you can use extremely oversimplified rivers if you only care about the flop strategy. You can use very basic postflop trees if all you care about is preflop.

More complex trees are better for database stuff like exploring aggregate reports, searching for overarching heuristics, or hand-history analyzers. A lot of strategies were only discovered because people explored weird sizes. e.g. flop overbets, oversized pf 3bets in the blinds, 10% river donks, etc.

Higher accuracy is best for comparing strategic options, like finding the best simplification for example. If your solution is too noisy then your margin of error becomes greater than the difference between the EV's. Higher accuracy has the added benefit of reduced noise in the strategy, so you get more pure actions and less mixing, which makes it easier to practice with.

I ran a lot of trees with scripting... 1% had noticeably inaccuracies (which I would call obvious "non-convergences", 0.5% was noticeably better but occasionally you can see the same/similar non-convergences, but this isn't usually a problem. Difference between solving to 1% to 0.5% is maybe not that big of a deal in terms of time/resources, but going from 0.5% to 0.25% and lower often takes twice as long with not as much noticeable differences (at least to me).

How do you identify obvious non-convergences?

Probably just arrogance speaking, should restate it as "probably non-convergences".

Weird mixes/lines. Stuff that is done like 5% that probably doesn't make much sense. Something similar to what you're seeing on that KQX board where you obv know you should bet a lot if you've ran many solves. Something like that but to a lesser degree.

What accuracy does GTO wizard solve to?

The latest solutions solve to 0.2% - 0.3% of the pot.
https://blog.gtowizard.com/status-an...our-solutions/


I suppose it comes with experience. One of the easiest ways to identify non-convergence for me involves just looking at the EV of mixed actions. For example, if a hand is folding/raising/calling then theoretically all three actions should have the same EV. The difference between those EV's is a good indicator of the noise in the solution.

Depends on the type, but 0.2% to 0.4%

I think flop tends to converge relatively quickly, so 1% would not be bad if you’re just looking at the flop.

Obviously 8% is a bit much but maybe the turns and rivers get resolved on the fly? Which would make it a little less egregious.