Good question, It actually does basic algorithm for a couple of rounds and then does the full algorithm once, then basic a couple of times, then full once etc. The basic algorithm unlocks a lot of squares quickly, but it happens that doing it too much (even when it's not stuck) leads to jagged edges and thus very long frontiers (bad!). So actually the "smart" algorithm is run for 2 reasons: to solve ambiguous cases accurately when the basic algorithm is inevitably stuck, and to "smooth" the frontier/perimeter every now and then to keep it short.

Ok will do

Theres a couple of tricks! 0. Making random guesses at any uncertain juncture yields abysmal winrates, maybe 5%. 1. As explained in the video, when no mines are certain, we click on the least likely square to be a mine. (Sometimes, this isn't the best choice as there are cases that the least likely square doesn't reveal on average as much information as another slightly riskier square which would have given more clarity. But this is not implemented here.) This gets the bot to about 49-50%. 2. Not explained in the video, (but is in the code), to more accurately calculate probabilities, the total number of mines in the minefield actually make a difference. In general, perimeters with fewer mines are more probable, since they have more mines left over and therefore more arrangements of remaining mines in the remaining squares. More care needs to be taken with calculating the probability of an arrangement when there are multiple perimeters - the number of mines in each perimeter affect each other and needs to be taken care of. Doing this careful combinatorial work improves the accuracy by about 1.5-2%. In fact the average accuracy is about 52%, the video's 50% is a below-average run.

@@codingchannel4422 Nice!, I have actually implemented your method in C++ but was able to get only 30% accuracy. Is there a particular function for arranging the mines in the start? Currently I am using rand() to arrange the mines is it correct or the minesweeper board has a different arrangement?. And also the world record is around 40%! (PSEQ - D256 is the method name). You should publish your work as a research paper!

@@madhukartemba2987 Very nice paper you linked! I think for my accuracy it is using different rules to the paper. In mine it is mimicking windows minesweeper where the first move is guaranteed to expand (have 0 neighbours), not just safe. It does this by randomizing the grid until it finds one where the point you clicked can expand. Well done in implementing it in C++, For your implementation, I think (barring any other bugs) it will be of similar accuracy once you have the same rules about the initial click. The paper already references using the method of choosing a mine with least probability (not sure if they calculated them accurately or not) and they added heuristics on top of that, no doubt it is going to be more accurate than mine with the same rules. I think that paper is interesting! The entropy heuristic is something I though of doing but alas didn't have time. I thnk it would be interesting to use deep neural nets for large unsolved areas (when the remaining game is still too big for an optimal game tree analysis). It did wonders in chess/go, minesweeper could be next :P The decisions of the neural net could be evaluated against all possible choices with monte carlo during training, where the remainder of the game would be done with a different algorithm (faster) or the DNN itself.

​@@codingchannel4422 Thank you coding channel for this video, it has helped me a lot in my project! I have also implemented the 0 neighbor part but still get accuracy around 35% (must be some other bug). Anyways thanks again for the video! Please check my implemented game here! madhukartemba.itch.io/minesweeper-ai Please reply if you find any obvious bugs/mistakes!

Oh awesome! Only just saw this, I'm checking it out that's cool!

Really appreciate your comment, and I'm glad you found this interesting! I have added my contact on my about section, it is maxwucoder@gmail.com. We can continue via email!

Guessing is also part of the fun.

You can't reach perfection like that. If you looked into how minesweeper is NP-complete, you'd know how far away from the tile you are considering you may actually have to check.