ClonesGame.com Forums

[rt]

One of the features we wanted to include in Clones was the concept of a "Boss Battle", where the player is essentially playing multiplayer vs. the computer. This turned out to be a huge problem and we had to settle for the current "Multiverse Match" CloneMaster battles in which a pre-recorded solution simulates head-to-head gameplay.

Creating functioning practical AI for this style of game might be impossible given today's hardware. Here is a paper that talks about a pathfinding algorithm for Lemmings: http://www.cs.nott.ac.uk/~gxk/papers/cec2004kts.pdf

The problem is even more complex in Clones because pathfinding is just one aspect of multiplayer, there would also have to be a sabotage algorithm and an active-defense algorithm all executing in real-time and computing successful strategies for situations being encountered for the first time. If there are any computer science students out there looking for a Ph. D. project to tackle then please let us know, a team effort would be welcomed

[minimac]

I'm afraid I can't help, but I know a couple of levels with AI trying to destroy your path:

DrillDude's Heart of Clone, RoboBrain's Atomize this and Roast Ed's Secret Garden. Pretend that these are boss battles for now, without the actual saving of any AI.

Also, I've recently created another multi-player level with AI trying to ruin it. I'll bring it out eventually.

[rt]

Yes those levels contain scripted actions for some non-controllable clones which gives a good illusion of AI. In fact, using a complex set of logic traps you could have some clones "react" to player actions, to a limited degree.

I'm much more interested in having full blown AI which would rival the best human multiplayer opponent. Maybe we can rent Deep Blue?

[minimac]

I'm much more interested in having full blown AI which would rival the best human multiplayer opponent. Maybe we can rent Deep Blue?

Good idea!

[CCX]

The recent IBM special for American game show Jeopardy!, featuring computer contestant Watson, reminds me of this thread again.

Here's a thought : see if you can get someone on the Watson research team at IBM addicted to multiplayer Clones. With any luck, maybe you can get IBM's 3rd big AI project to be a Clones-playing machine!

They might not even have to come up with a new name. After all, "Watson" is already eerily close to "Wattson"......

[CCX]

The problem is even more complex in Clones because pathfinding is just one aspect of multiplayer, there would also have to be a sabotage algorithm and an active-defense algorithm all executing in real-time and computing successful strategies for situations being encountered for the first time.

Instead of trying to design a general Clones AI capable of acting intelligently on any given level, maybe you will have more success by having a program that can learn from playbacks of human battles in existing multiplayer levels, and from those examples develop strategies specific to each level. It's doubtful you'll come up with something this way that can beat the best human players even on just those specific levels (after all, it took a few years for Watson to get to the level it was at with Jeopardy!, and then we still saw some obvious quirks during the televised Jeopardy! event), but maybe it'll be good enough for less ambitious purposes (eg. stand-in for human players to fill out multiplayer games on those specific levels, without having the AI be complete idiot). It's of note that in designing both Deep Blue and Watson, there was a lot of machine learning involved (feeding in past grandmaster games for Deep Blue, and past Jeopardy! questions/clues/categories for Watson).

[tom]

Instead of trying to design a general Clones AI capable of acting intelligently on any given level, maybe you will have more success by having a program that can learn from playbacks of human battles in existing multiplayer levels, and from those examples develop strategies specific to each level.

Using playback files to train the program (we need to come up with a code name for this computer controlled player!) with high level strategies for a given level is a good idea. For this to work, it would require the program to use machine learning techniques.

I feel that it's possible to make a program that can compete against human players, but the key it making it work is to be able to represent the level in a form that can be easily digested by a computer. Right now, a level is represented by individual pixels, but we humans do not think of it this way. Instead, we think of a level as made up of: walls, ramps, bridges, etc. So the computer needs to do the same thing.

It would take a lot of time and energy to make full fledged working AI for clones (IBM's Watson cost $25 million or more). But I was thinking, maybe we can start small and make a Clone multiplayer "assistant". The idea here is that a human player can enable the "assistant" and this would basically add a computer controlled player on their team. The assistant would then watch over the game and only participate when there is an obvious problem. For example, if clones are falling to their death and the human player doesn't do anything, then the assistant will either apply the puff morph or atomize morph to those clones. Similar heuristics can be applied in other situations.

Just a thought