Improve and extend microbe AI

[jjonj]

Features that the new AI should be able to handle:

• Different modes (predatory, plantlike, prey-like possibly)
• Surviving, seeking out necessary compounds and feeding on them
• Killing other microbes (predator mode), this involves using agent vacuoles effectively
• Fleeing from predators (prey-like mode, and possibly predators from larger predators)

More ideas are welcomed

[NickTheNick-zz]

I think what you said sums it up quite well. Some important traits I can think of that a microbe can exhibit varying behaviours in are:

• How much it is willing to put it's own security at risk for potential gain
• How much it explores the environment and tries to interact with things
• How much it prefers to stay near other species members

[ghost]

The levels of danger and such could be handled using variables (0 to 100?). If the danger level is low enough, it will ignore or attack. If the danger level is too high, it will flee. The threshold could be different for different species and in the case of a predator could also relate to the need of compounds (if the need is high, it will take a higher risk).

Searching for compounds could be done with a scanner function, but I'm not really sure how to code it. This could be done by having the function scan in a small radius from the microbe, and identifying all the compounds (heat, light, microbes?) within the radius. The microbe could then go to the closest concentration of the compound that the microbe needs most at that moment. Note:The radius shouldn't be too big, it would get unrealistic from a certain size on (also depending on the size of the microbe?).

How much it prefers to stay near members of its species could be handled by having 2 variables. Distance and type: distance: the distance between members of the species and type defines if it will stay near, neutral or will stay away. Distance isn't used when type is neutral. Distance means the furthest away a microbe will go away from its species when type is near. Distance means the closest a microbe will come to a member of its species when type is away. The microbe will ignore members of its species when type is neutral.

This is how I imagine the things that Nick stated will work.

[Moopli]

Some thoughts:

• FSMs are always a good starting point
• We'll need to be able to weigh goals by risk/reward
  • The main reason to make a goal weighting system is that it can be adaptive -- that's important, since the only way we'll be able to predict all possible microbe behaviours is by heavily limiting the scope of what microbes can do, which is bad.
• That covers the internals of a microbe, but what about communication with the environment?
  • What I'd prefer, for realism and some interesting side effects, is to use compound clouds for nearly everything:
  • For keeping friendly microbe together, we can use signal agents -- each signal agent can be marked internally with the source species, to abstract away the fact that different species will doubtlessly have their own signalling molecules. Then, not only can friendly microbes swim up the gradient to congregate, but predators can do the same to hunt -- it's a realistic consequence for heavy communication, after all.
  • For figuring out the distance from and direction to food sources, we simply read the compound concentrations around the microbe. But wait, you'll say -- what if there are no compounds to read around the microbe? Well, in that case, the microbe is screwed anyway -- realistically, there is absolutely no way it could know which way food is, so we could just leave it to do a Lévy Flight or something. Of course, since we'll likely never have a large enough empty space for a random walk to be worth the effort, we can just leave them to swim straight until they either find something or despawn.

Now to get in the nitty-gritty: for a risk-reward system to work, we need actual things to remember the risk/reward of. I suggest we use compounds for that -- since, if they're what we're using for environmental communication, they're the only option (and a good one too).

• Risk: The risk associated with a compound is pulled up or down based on how often chasing the compound results in harm. So, with time you get a steady decline, and whenever a microbe dies or suffers some other harm, the risk jumps up accordingly. Risk is measured species-wide, and is part of behavioural evolution.
• Reward: The reward associated with a compound is a function of the internal state of the specific organism -- depending on what its compound needs are at the moment, reward rises and falls. Reward might also rise for a compound the more there is available for it, but I haven't thought enough on that to decide if it's a good idea.

Whew, longer than I expected.

[jjonj]

Good notes, but would you suggest delaying this until we have compound clouds then?

Also I wrote some stuff regarding my experiments with OpenVDB to you on skype

[Moopli]

Mostly -- some basic framework could be figured out right now. The AI will likely no matter what need a few iterations to complete, from what I can tell; so having one (more, since what we currently have counts too kinda) iteration done before compound clouds wouldn't be too silly.

[Untrustedlife]

Ai has been improved and extended.