Forum › Forums › Discussion › Crafting efficiency vs real time efficiency
This topic contains 7 replies, has 4 voices, and was last updated by he_le 7 years ago.
-
AuthorPosts
-
October 28, 2017 at 4:53 am #5586
The crafting solver tends to favor actions such as hasty touch/flawless synthesis(buffed with maker’s mark) as these give progress/quality gains at no cost. While I can see how from a simulation perspective this makes sense, it also leads the crafting solver to use 40+ steps on basic items such as 40 durability ingots. This means that it takes 2-3 minutes to craft one item and is a overall loss in gil/min
Is there any way to avoid this besides taking the relevant skills off your actionbar? I’m not sure what algorithm you use for the AI simulation, but if it is anything like a genetic algorithm/function optimization, is there any plans to allow advanced users to specify weights?
October 28, 2017 at 11:14 am #5589October 28, 2017 at 6:45 pm #5590The algorithm we use is called State Space Search with Probabilistic Transition Matrices.
The weights are dictated by actions themselves, and they are hardcoded under several layers of AI. Extracting them into advanced settings would break low level optimizations, increase cache miss rate, and result in computation time burst up to several days for a single synthesis.
However we can adjust weights on development side, and for each new version of Miqobot we release a better version of AI.
The problem is that the current version of Miqobot still uses Heavensward AI for crafting which is not suited for Stormblood recipes. Crafting Solver will behave differently in Stormblood Upgrade Stage 2.
We would kindly ask you to wait until the next version, and then let us know whether Maker’s Mark behaviour still needs to be adjusted.October 28, 2017 at 8:43 pm #5591October 28, 2017 at 9:28 pm #5593Yes, a simple toggle on / off for each skill is planned to be added in Advanced Settings.
However this is exactly the same as taking unwanted abilities off your hotbar.The original question is about adjusting weights of the algorithm, which would allow Miqobot to identify whether a certain skill should or should not be used automatically. This feature is technically not possible outside of the source code.
October 29, 2017 at 4:43 am #5595Thanks for sharing, very interesting you use state space search. Guess I will wait for upgrade stage 2 and take actions off my bars in the meantime
As an aside, apologies if this is asking too much information about your design, but what is the reasoning behind using state space search compared to tree search or GP for optimizing the rotation? My prior experience with similar game problem has involved tree search(MCTS for ex) and am surprised to find state space search being used here. Is this due to memory savings of state space search?
October 29, 2017 at 6:00 am #5598Crafting Solver does not use any rotations.
It calculates the required crafting ability on the fly by analyzing all possible outcomes of every possible decision. We use about 20 different methods for research, analysis, and space clusterization, and Monte Carlo tree search is of course one of them.While we can provide more details of this process, the amount of information would probably fit a PhD thesis and would not be very useful to the commumnity. All practical results of our research are being shared in the Upgrade Stage 2 thread.
Thank you for your interest very much!
October 30, 2017 at 8:49 am #5606 -
AuthorPosts
You must be logged in to reply to this topic.