Najbolj popoln in pregleden TV-spored za Planet TV, POP TV, Kanal A, SLO 1, SLO 2, HBO in druge TV programe. Kategorizirani dnevni TV-sporedi z opisi in ocenami oddaj. And Analysis System DeCA Defense Commissary Agency DEFCON defense. Drift error de minimax minimax drift error DeMS deployment management system. DEP Delayed Entry Program; deployed DEP&S Drug Enforcement Plans and.
Over the past month I have been working on a way to implement a better AI in Melee. I got really interested in this project after seeing how useful the teching/DI'ing was in 20XX and I wanted to see if I could do a complete overhaul of the AI.
![Minimax Program Za Decu Minimax Program Za Decu](/uploads/1/2/4/0/124070213/822910684.png)
I finally have enough code written to show something, so I made a quick video demonstrating what the AI has so far, which right now is just recovery. I wanted to start with recovery because I think edgeguarding is still one of the hardest things to solo practice and I think this could really help. Main Goal: The main goal of the project is to create a better AI for the purposes of TRAINING. I'm not trying to make an unbeatable, TAS level computer player at all, in fact its nearly the complete opposite. I'm really trying to focus on implementing the most human-like AI possible, since anything else would just enforce bad habits. For that reason things like neutral game and combo game (AI comboing human) are going to be extremely difficult, if not impossible.
However, I do think it is reasonable to make an AI that plays defense in a human-like way, and so these are the main goals I have right now for the project: 1) recovery for all characters on all stages 2) teching 3) DI'ing (human-like as possible) 4) attempts to break out of combos. sheilding after tech - OOS option. jump / throw out hitbox right out of hitstun. airdodge.
etc. #2,3 are already in 20XX, but since this approach to the AI allows for more control we may be able to get more human-like patterns. As for #4 I think alot of progress was made in 20XX with the jumping out of hitstun, but I think this appraoch to the AI could improve on it further. And finally, #1 recovery, will be the main focus of this project since edgeguarding is so important and right now there are no tools to practice it solo. Current Status: The recovery logic is 80% written for fox and now just requires more testing. Teching and DI'ing should be easy to add and I'll do that next.
Nothing has been done towards #4. However, out of the month I've been seriously working on this, only the last 3 days were spent on the AI. The rest of the time was spent on the framework surrounding the control of P2. Now that the foundation has been established, I am optimistic about how quickly I can write the code for the AI. Future Possibilites: I am also thinking of ways people can use the framework I've made to control P2 and get data from the game to make decisions. One possibility would be to create training setups similiar to Falco SH lasering, but instead of rapid fire I could make Falco dash dance and shoot at random heights/timing. Furthermore, I could organize my code and release a library so that other people can play around with this stuff as well.
How It Works: So now for how this atually works, essentially I modified the dolphin source code (which has read/write access to all of melee's memory) so that every frame dolphin writes data from important memory addresses into a block of shared memory with my program. That way my program has lag-less access to all the info in the game.
In addition, my program writes the data for the P2 controller into shared memory and dolphin then writes this to a free chunk of melee's memory. Then I inserted assembly code through a C2 gecko code, which copies the controller data from the reserved chunk of memory into the live P2 controller addresses at the right time. And if you've been following along closely you probably realize the biggest downside to this approach. Right now there is no way I see this being ported to console, i.e.
![Decu Decu](http://i.ytimg.com/vi/IWb4dwzDFM4/0.jpg)
You need to be running dolphin on PC. There have been posts on smashboards from someone (i think SaveState) who came up with a way to compile C code into asm code for melee, so that gives me hope that a simplified version of the AI could be put on the console, but I doubt the full program could ever be ported. Edit: (link for C code - ASM) The purpose of this post is two-sided: For people interested in helping with development: I would love it if some people were interested in helping with the development of the project.
There is simply no way I could accomplish all of the goals by myself. If you have experience coding (paricularly C) and are intersted in helping please email me at cause I could use the help. Even if you aren't a programmer, one of the biggest challenges for me is that I'm relatively new to Melee so I don't know the recovery tree for all the characters (in fact I only know it for spacies, falcon, marth, shiek). So it would incredibly helpful if some people could write out the recovery logic for characters (including certain mix-ups and how having a wall changes things) For people interested in using this program: Right now the program is HEAVILY in the development stage, and is very very not user friendly (for starters you need to be running Dolphin in Linux).
I really would like to know the interest level among people who want to use this for practice, since that will determine if I focus on further development or spend the next few weeks getting this ready to release a user friendly version. I really believe this could be useful to many people, and even if not its been a hell of a lot of fun for me.
Edit: Smashboards post: GitHub:. This is really cool. A couple suggestions on Fox's AI logic:. It seemed like Fox was always going for either up b or side b, even when it could double jump back on stage. If you could somehow program it such that it can recognize when double jumping to the ledge or the stage is possible, then you can incorporate the mixup of either sweet spotting the ledge with its double jump, double jumping back with a drill or another attack, or double jump airdodging onto stage. It seemed like Fox was side-b'ing to the ledge a little bit high, and not quite getting perfect wavelands back on stage.
I'd imagine those are pretty easy fixes, so you should consider doing that. Awesome work you're doing. Recovery definitely seems like one of the hardest parts of this project, so good luck!