Rank: Newbie
Groups: Registered
Posts: 2
Thanks: 1 times
Was thanked: 1 time(s) in 1 post(s)
|
Hi, I m a new user of your amazing tool (thank you for doing this :) ) and I post here to suggest an idea. I read that some bots working with a special algorithm could learn from their previous mistakes. It seems very difficult to code but I think it could be great (if it's possible) to add this functionality to the Hearthranger's AI. For example if a player play a card and the bot doesn't clean it but next turn the player buff the card and kill me, the bot will "know" that it have to clean this card because there's a probability that the same move come up again. The bot could "watch and learn" a lot of games and with all those datas it could deduce the probability of a certain move and the best move to do to counter it.
I'm sorry if there are spelling mistakes i'm a french people.
|
|
1 user thanked EtAk0 for this useful post.
|
|
|
|
Rank: Administration
Groups: Administrators
Posts: 4,847
Thanks: 805 times
Was thanked: 4539 time(s) in 1637 post(s)
|
Till now there is no effective way to 'coach' bot, it's failed in other bots too which provides such function.
If any research guys with deep learning knowledge who can link bot's a.i. to real machine learning network, it has chance to learn from human beings.
|
|
|
|
|
|
 Rank: Advanced Member Groups: Registered
Posts: 267
Thanks: 35 times
Was thanked: 26 time(s) in 23 post(s)
|
This is a really ambitious suggestion
|
|
|
|
|
|
Rank: Advanced Member Groups: Registered
Posts: 31
Location: EU Server
Thanks: 3 times
Was thanked: 1 time(s) in 1 post(s)
|
I was thinking about this also, a self learning bot. It's definitely doable. But would require great AI skills from the dev.
A bot that learns the type of decks used by players based on the cards played and mulligans ecc... it would then make a database and try to understand what kind of deck the oponent is using based on that databse, to play accordingly and increasing the chances of winning and even better to change the deck cards to beat them.
Seems like a lot of work tho, and if such a thing would be real Blizzard would probably spend more resources on chasing botters.
|
|
|
|
|
|
Rank: Advanced Member Groups: Registered
Posts: 40
Thanks: 13 times
Was thanked: 17 time(s) in 15 post(s)
|
It is doable but why waste the time for something like that? This would be equivalent to a thesis on artificial neural networks.
What would be more realistic is the ability to "inject" rules to the AI in the same way as mulligan and combo work. You could "program" a bot to use a specific meta deck with specific rules vs all kinds of other meta decks. Otherwise people can use silverfish AI as their base to actually program their own AI specific to their deck. This requires a significantly lower amount of time and is less complex. (Could be done in a few weeks probably)
Do not forget Elie Bursztein who actually programmed a machine learning tool to find undervalued hearthstone cards and to predict what would be played next with a very high certainty. And what followed next. Any creator of such advanced tools will be the primary target of the legal department of Blizzard.
All in all its a high risk project with a low reward.
|
|
1 user thanked contagious86 for this useful post.
|
|
|
|
Rank: Member
Groups: Registered
Posts: 29
Thanks: 9 times
Was thanked: 3 time(s) in 3 post(s)
|
I think you're right. I looked into Elie Bursztein because of your post. And I have to say a bot that does what OP wants would be high risk and low reward. However, I think a AI that looks for patterns in the cards that the opponent plays and even has data on the meta (user imputed as well), and then for that AI to feed that info the bot to slightly alter the way the plays to make better choices. The bot will consider the probability of a secret and try to activate it or play around it and get the best outcome. I think that would be worth the effort.
|
|
|
|
|
|
Rank: Advanced Member Groups: Registered
Posts: 31
Location: EU Server
Thanks: 3 times
Was thanked: 1 time(s) in 1 post(s)
|
Well actually there is a AI company co-founded by Elon Musk. They are working on a AI that is primarily focusing on advancing machine learning, which is the technology that enables computers to learn how to complete tasks through experience. And the AI will be open source. The company is called OpenAI.
I doubt we see any self learning bot in less than 4 years in Hearthstone tho. But when that day will open, online games will have a serious problem with bots...
|
|
|
|
|
|
Rank: Advanced Member Groups: Registered
Posts: 40
Thanks: 13 times
Was thanked: 17 time(s) in 15 post(s)
|
Its funny to even mention Elon Musk here.
But yeah it is hard for someone to find the motivation for something like this, even though its not that hard but still challenging.
|
|
|
|
|
|
Rank: Advanced Member Groups: Registered
Posts: 116
Location: Denmark
Thanks: 10 times
Was thanked: 32 time(s) in 25 post(s)
|
I read an article recently actually about machine learning software by Google: http://www.theverge.com/...ning-tensorflow-upgrade
the software is open source: https://www.tensorflow.org/
Maybe with the right coding it can actually be made possible, would be crazy to gather information serverside with new and possible combos by learning from opponents over time (and minimize mistakes done by the bot potentially) and share it across all HR bots. I know it might never be done or is a huge task which takes much effort so its not anytime soon (I guess) but the potential is great.
|
|
|
|
|
|
Rank: Newbie
Groups: Registered
Posts: 4
Thanks: 1 times
Was thanked: 2 time(s) in 1 post(s)
|
Not all machine learning implementations would have to be that intense.
A simple evolutionary algorithm can be used to select out over 'generations' for the AI setting that has the best winrate. Then continue to adapt by mutation while always selecting out for best winrate.
It's effectively a way to let the computer figure out which AI settings would be best for a given deck based on
total winrate.
I'm working on this now.
|
|
2 users thanked Crixxus for this useful post.
|
|
|
|
Forum Jump
You cannot post new topics in this forum.
You cannot reply to topics in this forum.
You cannot delete your posts in this forum.
You cannot edit your posts in this forum.
You cannot create polls in this forum.
You cannot vote in polls in this forum.
