Hacker News Reader

Comment by Dotnaught | Hacker News Reader

If you’re modeling an intelligent pursuer, the algorithm should anticipate future travel rather than just following.

In games the goal of the developer isn't necessarily to have intelligent agents, but to have agents that are fun to play against. Sometimes the goal is to have them be intelligent, but sometimes having them behave in "dumb" but predictable ways makes the game overall more fun.

This is the reason usually given for using relatively simple techniques like behavior trees, but, anecdotally, I find that the biggest let down in most games is that the AI is so dumb that it 1) is immersion breaking, and, 2) get same-y and boring very very quickly.

Yes. It depends on what the game is trying to achieve.

Subset Games, the makers of FTL and 'Into the Breach' have talked about this extensively. 'Into the Breach' deliberately has the enemies telegraph their plans one turn ahead of time, and the game is all about interfering with those plans.

The rest of the game's design carefully reinforces the message that the enemy units are not intelligent. The backstory has them as basically oversized insects.

If you have a game that pretends to give you realistic human antagonists, but they behave mechanically dumb and predictable, that breaks immersion like you suggest.

One big problem is that having very smart AI that's purely there to oppose you in a zero sum game ain't fun to play against for most people. The handicaps a modern chess or Go engine would have to give you a normal human for a fair fight are ludicrous. And people seldom want fair fights in their games. They want a feeling of accomplishment, but without actually putting in all that much work.

Even hardcore games like XCom cheat in your favour behind the scenes.

There are at least two ways out of this while still avoiding the boring repetition:

- carefully make the NPC make believable human-like (or animal-like) mistakes, instead of easily exploitable repetitive mistakes - give the NPC goals that are in conflict with the player, but not 100% so.

A silly example of the second option:

Take a game like Thief that's all about sneaking around stealthily and stealing stuff. Now realistically, most of the guards are just hired goons. They don't want to die, but they don't particularly care about protecting the place. They do care about being seen doing their job, so they don't get fired.

So your job as a player could be, in addition to staying unseen, to provide plausible distractions and reasons for the guards not too investigate to closely.

Higher ranked guards, and owners, would be under higher pressure to perform and won't get away with excuses. So they would be more alert.

Using the same trick over and over again would lower it's effectivity: guards can't plausible claim to their higher ups to have been tricked again and again.

If you are starting to become aggressive to a guard, or he learns that you called one of his friends, the guard's priorities will change towards more self-preservation.

A pacifist run might even earn you respect and admiration from the lower level guards. Just like cat burglars are often admired in real life.

Seen a bit more abstract, the game now becomes one of three factions: the thief (that's you), the low level guards and their employers. All with partially overlapping, partially conflicting goals.

You can throw insurance companies into the mix, if you want to make it even more complicated.

Thanks to the non-zero sum nature of the partial conflict, you can crank up how smart everyone acts, without overwhelming the player:

Eg smarter guards might figure out a way to de-escalate that still looks like a plausible and even courageous move to their employers.

I think you've hit the nail on the head. I absolutely agree that "dumb" AI can be explained away and integrated into a game's design and story in ways that make it much more interesting and believable than typical bad AI on its own. I also agree that by providing a flexible and interesting enough scenario, with conflicting goals and motivations, you can improve the AI in ways that are both noticeable to the player and add extra layers of interesting gameplay.

As an aside, unrelated to your reply, I just remembered another common excuse, which is that the enemies only exist for X seconds before the player shoots them, so effort on smarts would be wasted on them. For some games, I think its for sure the case, but I also feel that in many cases the enemy only exists for a few seconds because they are dumb and uninteresting.

All agreed.

Both 'dumb' and 'interesting' AI can be useful in a game, it all depends on your game's design.

To give another silly example: Tetris by default has very 'dumb' AI that just gives you pieces at random. You could imagine variants of Tetris with more interesting piece selection.

For example, an AI that makes pieces as unhelpful as possible while keeping their distribution statistically indistinguishable from true random selection.

Or there was an inversion of the 2048 game, where an AI plays the normal game, and your task is to give them unhelpful numbers.

> it can check if it can see any of these past positions, then follow them to the player. Since this is similar to how a dog would track something, we'll call it a scent trail.

This is how a scent dog tracks prey it cannot see.

I recommend the author hang out at the dog park and watch other kinds of working dogs, like retrievers and herd dogs. Very, very different behavior.

In fact I've heard that dogs can make a pretty nearly accurate guess about shortest path when dealing with terrain that has two different speeds. If, for instance, your dog is trying to get to you in the water and you are diagonal to his position, they will use two angles to triangulate on you, staying on the sand longer (but not until perpendicular with you) to minimize travel time.

One of my dogs definitely uses the Wayne Gretzky approach (go where the puck will be) but sometimes arrives before her target, resulting in collisions or acrobatics.

Interesting observation!

I wouldn't fault the authors too much, they just looked for an easy to remember name and the metaphor holds up well enough for that.

I may have come across as a bit condescending there.

Really, if you wanna do path finding, a dog park will give you tons and tons of food for thought.

Some dogs are very clever. Others dumb as a post.

Yes, lots of things to learn from dogs.

Insects and other animals are also good for inspiration. Even microbes!

There are some microbes that essentially follow this small program to find sugars:

- swim forward - wiggle around randomly - repeat

When the concentration of sugars is increasing, spend more time swimming forward. Otherwise, spend less time swimming between wiggles.

In a game, you would have the player leave a scent trail, and some of the other critters could behave like above.

Yes. But for simple enemies, the algorithm suggested gives interesting behaviour, and plenty of ways for players to manipulate it.

The latter is useful to open up extra gameplay.

You might be able to achieve something like that by storing not only the position of the player in the "scent" but also its direction and speed. Then you can simulate this ghost image of the player that moves linearly in this direction and that the enemy targets.

Of course at this point you probably end up with something that's actually more CPU-intensive than a naive A* so it can't be used as an optimization. I think the cleverness of TFA is that it combines both a performance optimization and makes the enemies look smarter than omniscient bots doing A* perfectly every time.

Reply on news.ycombinator.com