1 of 11

Deep Dive: AI

2 of 11

Deep Dive: AI - Movie + Dinner

3 of 11

Deep Dive: AI - Movie + Dinner

pick movie
pick restaurant nearby
buy tickets
make reservations
drive to restaurant
eat
drive to show
watch show
drive home

Right, so first before you go see the movie, you have to pick the movie. Based on the movie you pick, that narrows down the possible theaters you can go to and the times that you'll be seeing it. This in turn narrows down the restaurants available to you. You pick a restaurant based on these constraints. Then comes the verification step: can you in fact, in advance, through your phone, buy tickets for this movie, and make reservations for the restaurant? If you can neither buy tickets or get a reservation, you may have to redo one or both of the picking steps--either pick a different movie or pick a different restaurant. Many permutations.

Once you have decided on the movie and restaurant, you can then execute your evening--drive to the restaurant, eat, drive to the movie, watch it, and go home.

4 of 11

Deep Dive: AI - Action Stages

pick movie
pick restaurant nearby
buy tickets
make reservations
drive to restaurant
eat
drive to show
watch show
drive home

Planning

Acting

Cleanup

5 of 11

Deep Dive: AI - Action Individuality

pick movie
buy tickets
drive to show
watch show

pick restaurant nearby
make reservations
drive to restaurant
eat

drive home

Movie

Restaurant

Return

Well, yes we could, but planning/action isn’t the only way to divide up the steps. Here… you can see that this is another way we’d divide up/categorize the steps. We can also separate everything involved with the movie, from everything involved with the restaurant, from everything involved with the driving home. Can anyone think which might be better/more useful and why?

Well, as any good engineer knows, you also usually want to write your code in such a way that like-functions/behaviors stay together, so that you can re-use your code. What if I wanted to write an AI in which I visited only a restaurant, or only went to a movie? What if I wanted to write it so that I went to a restaurant and then ice skating? Dividing the actions up this way would allow me to do that, at the expense of having actions that can be easily interleaved and dependent on each other.

So what's the solution? How do I get interdependence but also re-usabiity?

6 of 11

Deep Dive: AI - Nest Like-Units

do movie + dinner

pick movie
eat dinner nearby (input: location [of movie])

find nearby restaurant
reserve restaurant
go to restaurant
eat

watch movie (input: current location, movie)

buy tickets
go to movie
watch movie

drive home

7 of 11

Deep Dive: AI - Run Planning First

do movie + dinner

pick movie
eat dinner nearby (input: location [of movie])

find nearby restaurant
reserve restaurant
go to restaurant
eat

watch movie (input: current location, movie)

buy tickets
go to movie
watch movie

drive home

8 of 11

Deep Dive: AI - Then All Acting

do movie + dinner

pick movie
eat dinner nearby (input: location [of movie])

find nearby restaurant
reserve restaurant
go to restaurant
eat

watch movie (input: current location, movie)

buy tickets
go to movie
watch movie

drive home

9 of 11

Deep Dive: AI - Remix

do dancing + dinner ()

pick club
eat dinner nearby (input: location [of club])

find nearby restaurant
reserve restaurant
go to restaurant
eat

go dancing (input: current location, club)

drive home

10 of 11

Deep Dive: AI - In the game

And this is what it looks like in the game. The dev version of the game has an AI inspector showing all the things that could be running on a person at a given time.

At the top you can see that though the person hungry, she’s running idle->despair (current actions are in green) because simultaneously, she’s trying to run eat, but it’s getting stuck in the thinking/planning part: she can’t find any food to pick up. Since she can’t find any food in follow path, she can’t reserve it and the running to the food will never happen.

--

(It’s a little hard to read because vertical alignment means both ‘peer actions--all the equivalent things I could be doing’, and sequential actions within an action, or all the things I need to do to make this happen, but let’s just focus on one action at a time.)

11 of 11

Deep Dive: AI - In the game, working

Once there is food in the world, she is able to locate and reserve it, and the planning section (thinking section) finishes, allowing the eating action to actually run. (In our dinner/movie analogy, she’s found the tickets and the reservation and is now eating/watching the movie.) The green means that the planning stage for the actions mentioned action have succeeded, and the AI can progress.

As you can see, actions have an incredibly fine amount of detail. The reason I used the dinner/movie analogy is because it is ONLY at an hour-long activity level that humans start to think that we need to break our behavior down, and plan everything in advance. We actually do the same when we’re trying to get a seat at a buffet, but it’s a much more automatic process. In Stonehearth, however, and all similar kinds of programming problems, the AI has to handle each little interaction as if each were a big deal.