"Programmable" AI that learns from a user

[meowreka]

In the some of the games in the Unarmored Core series of games, the player can "program" an AI by fighting in battles with the AI.

What I mean is, the AI would "watch" the user play games, and over time compile enough data to form a strategy. It "learns" how the user plays and adopts that style.

This wouldn't be terribly difficult to implement, and would be incredibly user friendly. A player could just tick a box at the beginning of each match to choose whether or not the AI will use the data from this match.

[johndh]

http://www.wesnoth.org/forum/viewtopic.php?f=12&t=10152

4. Give clear reasons why you think your idea should be implemented

Bad: "Draugs should have an extra attack."
Better: "The (level 3) Draug stats are only a very slight increase from the (level 2) Revenant's stats. In play I often hope that the Revenant won't level up because the increase in power isn't worth the increase in cost (+1 gold per turn). I propose adding an additional attack and +10hp to the Draug to justify the additional upkeep cost."

To quote Dave: "Any new feature is by default bad. It's bad because it's more we have to write, and more importantly, more we have to maintain and test. We will only implement a feature if it solves some problem that we think is worse than the cost of adding the feature. If you don't tell us about this problem, then we might not notice it for ourselves, and we'll see your idea simply as being 'bad bad bad'."

5. Clearly state how your idea will work in practice

Be specific. Numerous threads containing promising ideas have died because the posters weren't quite clear on what was being discussed.

Bad: "Mages should have an aura that damages attackers".
Good: "I propose adding a defensive flame aura to red mages. After each successful strike an opponent makes against the mage, 2 points of fire damage is inflicted on that opponent. The fire attack would always hit, but would be affected by a unit's fire resistance as per normal. The aura operates in addition to the mage's normal melee retaliation (staff)".

1. How do you propose the AI would learn?
2. What behaviors would it look for?
3. How would it change its behavior in light of the information it gathers?
4. Why is this feature needed?

[meowreka]

look at how uptight you can be, isn't it adorable

Nobody's going to even attempt to expound upon this idea because I didn't submit it in that format huh

well I won't either.

[Iris]

Sad, but you are the poster, not us.

[Daxion]

One of the main problems with an AI is, that it is very hard to exactly design an algorythm that does something that's easy for us humans.
We all know that the AI in Wesnoth could be improved, however it is hard to figure out how. So in that respect AI is the topic on which you should state how a proposed change should work.
Simon Mith did a very well proposition on that ( http://forums.wesnoth.org/viewtopic.php?f=12&t=30788 )

[zookeeper]

This wouldn't be terribly difficult to implement

---
------

---
------
---------
------
---

------
------
------
------
------
---


---
---

-
---

[ancestral]

I think meowreka should take a stab at it

In actuality, this would be an extremely challenging feature to implement. Possibly it could be a Google Summer of Code project. I don't know how you would tailor to or even recognize certain patterns of behavior, and what that behavior would be. Ideally you could take weights of all these different factors, but even still, coming up with the right weights could be Herculean at best.

Probably the best bet is to start with a foundation where this idea could even be possible. Much like having a computer play itself at chess to get better, one might be able to develop a heuristic through the AI fighting battles against itself, utilizing a genetic algorithm to evolve the AI and further optimize it. (If this could be done successfully, the results could be fantastic.)

I'm no AI expert; I only took one class in Computer Science, but yeah, someone else likely has more experience and knowledge in this area than I do and might be able to put in their two cents. In short, it's possible, but sometimes you need intermediate steps before you can get from one point to another.

[thespaceinvader]

---
------

---
------
---------
------
---

------
------
------
------
------
---


---
---

-
---

QFT. Writing a good AI is probably the most difficult coding task in the GAME. Writing a LEARNING AI is an order of magnitude beyond writing 'just' a good AI.

If it is, as you say, not so difficult, how about you do it, and stop being sarcastic to people who are trying to help you behave sennsibly?

Locked.