Maldades
Joined: Aug 27, 2014
|
 Posted:
Nov 19, 2021 - 21:37 |
 
|
Hello!
There is this annual competition on Blood Bowl bots based on Machine Learning, which is called BotBowl and is three years old and running.Unlike ginourmous SC2 or DOTA2 bots, this competition has a restriction of the bot being able to work on an average laptop, so maybe it could be of interest to people who use or admin this webpage.
You might think that, as other "more serious games" like Chess or GO have been mastered, BB is not very interesting, but you would be wrong. A measure of the complexity of a game is the "branching factor", which describes the available actions between which a player must choose in every turn. While Chess has an average branching factor of ~30 and GO's is of ~300, BB has been calculated to be around 10^50.
This simply means that old techniques have trouble working here, because trying to make a bot think through 10âµâ° decisions would surely lead to the first robot strike  .
This year's winner, MimicBot, first trained to imitate the actions of a scripted bot, and then was let free to explore and take its own decisions. After a week of training, it ended beating the scripted bot quite regularly and learning stuff the scripted bot was just not programmed to do.
It plays more or less like a novice BB player...
You can see it here in BB Youtuber Andy Davo's channel:
(youtube.com/) watch?v=uJ9txQwY4qA
It is easy to imagine that, if we subtituted the scripted bot's games by those of proficient humans, we would be able to archieve much better results, accelerating the arrival of a bot that is interesting to play against.
If you would be so kind to allow me to massively download games, I would share them with the BotBowl community, minimizing any disturbance this could cause to your servers .
So... yeah.
Whatever you decide, thank you for this awesome webpage! |
|
|
Kinks
Joined: Feb 28, 2007
|
| Posted:
Nov 19, 2021 - 21:52 |
|
Part of me thinks this is very cool, the other part is concerned this is the start of Skynet. |
_________________
Better lucky than good |
|
Sp00keh
Joined: Dec 06, 2011
|
| Posted:
Nov 19, 2021 - 21:53 |
|
Potentially very cool, I've previously thought about BB bots and how a lot of it is just weighting probabilities but a lot of it is actual strategy so therefore was very difficult for bots, and yea the huge amount of potential actions
I understand why you want the asset (2796769 matches played!) but not sure on your confidence you can extract useful behaviour from it
why's the website broken? https://bot-bowl.com/ |
Last edited by Sp00keh on Nov 19, 2021; edited 3 times in total |
|
Erik
Joined: Aug 02, 2003
|
| Posted:
Nov 19, 2021 - 21:54 |
|
Wow! That is awesome!
Is there any way us regular players can try and play vs the bots? Would love to try...
On topic hope you get what you need from the site!  |
|
|
Maldades
Joined: Aug 27, 2014
|
| Posted:
Nov 19, 2021 - 22:18 |
|
Search the webpage on google, its ont github, I cannot post external links...
I have to warn you that training neural networks is not as much fun as playing or watching bloodbowl, though 
Currently there are no available bots from the competition.
I think it's on the way...however, it's much more fun to play against humans, honestly... the bots are not interesting because of their tactical choices in BB...for the moment!
@Kinks: Yes, for the moment we are training them to have the board as input and winning the game as the reward signal that guides which moves should be considered as good, but giving them total control over military assets is the next planned step. We train our drones on pictures of people with moustaches and single eyebrows and hope for the best. |
|
|
MattDakka
Joined: Oct 09, 2007
|
| Posted:
Nov 19, 2021 - 22:25 |
|
I have a question: is it possible to have a bot playing vs a bot (instead of programming a bot with script) games to develop an AI?
Because, as far as I know, the AIs nowadays are not anymore programmed in the traditional way (like, for example, the very early Chess AIs), but with deep learning and euristic approach.
In simple words, an AI plays vs itself (or another AI) thousands of games in a short span of time (a thing not possbile to do with human coaches) and then, after thousands and thousands of games and experience, the AI improves its tactics by trials and errors and gets even and even better (which is how humans do: humans are not programmed to deal with problems, they learn by experience).
So, is it a possible scenario for BB AI or is it still sci-fi?
By the way, BB, compared to Chess and Go, has the extra complexity of the different rosters and skill builds. It's harder to train an AI to pick the right player skills, I guess. |
|
|
Grod
Joined: Sep 30, 2003
|
| Posted:
Nov 19, 2021 - 22:40 |
|
Yes Bot vs Bot exists, thats how the AI competition was run. There is a GitHub project which you can download and run and set up bot vs bot or human vs bot. Its also possible to run it as a server client model (like FUMBBL), but noone has done that yet in a public way, but you can set it up for yourself easily enough. For anyone that ever played against the GrodBot in Java, I ported it across into this format (the GrodBot won the first BowlBot and was runner up in the second and third BowlBots). For anyone who is interested, the GrodBot is a Scripted Bot, but lots of room remains to implement it further (its a bit rough), if you know Python Code and want to give it a try, but are daunted by full on M-L. Still, I am as excited as anyone when fully self-learning Bots start to dominate!
Some Links (the first is the bowbowl project, the second is a dropbox link to the GrodBot and the third is the original Java client I made which is still playable)
https://github.com/njustesen/botbowl
https://www.dropbox.com/s/7691w4urazkgn60/grodbot.py?dl=0
https://fumbbl.com/p/blog&c=Grod&id=14742 |
_________________
I am so clever that sometimes I don't understand a single word of what I am saying.
Oscar Wilde |
|
|
|
| |
Are you E.L.F. enoug...
Search
