Alex at AIGameDev: GOAP is dead

[amadmonk]

This Year in Game AI: Analysis, Trends from 2010 and Predictions for 2011*—* AiGameDev.com

In particular, Avalanche switched away from GOAP after JUST CAUSE 2, DEMIGOD's hybrid planner wasn't reused in SUPREME COMMANDER 2 at Gas Powered Games (it's now a neural network, of all things) and METRO 2033's original STRIPS implementation was replaced with a prioritized list of complex hand-crafted behaviors. The many issues with STRIPS are well documented, in particular that it requires you to add unnecessary variables to the world representation just to make it understand simple concepts, and that the only other way you really have to control the behavior is using weights & heuristics which also affect performance tremendously. I call this "baby-sitting" the planner, and it's a tricky balancing job to say the least.

Academia has long discarded such planners in favor of hierarchical ones, particularly when performance matters. Obviously for games it does too, but you can also control hierarchical planners much better thanks to their top-down task decomposition mechanisms. If you're thinking to yourself, Guerrilla Games saw this coming and were a step ahead with KILLZONE 2, then you're absolutely right (see this panel discussion at the Paris Game/AI Conference 2009). However, while GOAP has turned out to be a dead end, it's not necessarily the case that HTN planners will shine from now on. Unfortunately for planners, there's too much similarity between behavior trees and HTN planners, so it's much harder to justify the extra implementation time during development and computational effort at runtime.

^this. Basically, while the idea of planners sounds good -- feed in inputs and world state on one end and get out a viable action plan on the other -- the reality is that you spend all your time "babysitting the planner" as Alex says. It's like evolutionary algorithms -- just add water and your code "learns" and gets "better." Sadly, the reality never quite matches up to the hype, and thus "scripted" AI like BT's still rule the AAA game world (and we bot writers would do well to learn from the "big boys").

Edit: Erps... meant that for the elite forum. But I guess it's generally useful to bot writers, so I'll leave it here.

Edit 2: interesting, Alex also says that flow field navigation/pathfinding is taking off, which is something I anticipated a while back with my somewhat-breathless post on "magnetic field" navigation over in elite. Obviously I used the wrong terminology, but the idea of using field representations and (to some degree) letting the terrain "navigate itself" is very useful. In point of quibbling, I don't believe that flow field navigation is truly possible (due to the insoluble problem of local minima), but flow field steering is definitely possible, and a very useful tool.

[Sychotix]

Alot of bot writers DO use behavior trees btw =P

[amadmonk]

Well yeah, BT is probably the most common flavor of scripted behavior around, these days.

This was in response to a discussion I was having with some bot writers around here about using a GOAP for global planning. I was trying to express why planners were just overkill for the kind of AI that you need in a WoW bot, but I think I was making a hash of it. Alex, as usual, expresses it fairly clearly both in the fact that AAA game studios are ditching GOAP's (HTN's are still hanging on, barely) and in his statement that GOAP's require too much babysitting. BT's are much friendlier to build and tune, probably because they are completely parallel to general coding structures.

[namreeb]

I am self-centeredly (is that a word?) assuming I am one of the bot writers you were referring to, and in my defense I would just like to say that I had no real intention of using a planner to do the actual bot control. More as a tool to run offline to generate a grinding/leveling profile. I think it still has potential at a higher level to determine what quests and their corresponding objectives to perform, which to skip, and what order to do them in.

Edit: I forgot the obligatory "SO THERE! "

Edit 2: In seriousness, what I actually had envisioned is probably just a glorified A* once I can translate potential actions into some approximate completion time (to compare the efficiency of grinding in location x to completing quest y).. which may or may not qualify as a planner.

Necro edit 3: I just noticed Alex from aigamedev.com emailed me asking me to add this to the discussion:

"""
Behavior trees are not actually "scripted" in the traditional sense.
BTs do contain some specific sequences that are hand-crafted by
designers, which make them easier to use than STRIPS-planners, but they
offer equally adaptive behavior. The BT is constantly evaluating
options, making decisions, dealing with errors, and monitoring those
sequences to make sure nothing breaks the illusion. This is actually
what makes BTs such an interesting improvement of traditional scripts.
"""

[amadmonk]

Actually it wasn't you :P

Probably with the quest ordering issue you'd do better with a simple energy metric, except for the issue of quest rewards (which is the only part of quest planning that lends itself towards actual planning as opposed to just "pick the next best quest").

[Cypher]

I am self-centeredly (is that a word?) assuming I am one of the bot writers you were referring to, and in my defense I would just like to say that I had no real intention of using a planner to do the actual bot control. More as a tool to run offline to generate a grinding/leveling profile. I think it still has potential at a higher level to determine what quests and their corresponding objectives to perform, which to skip, and what order to do them in.

Edit: I forgot the obligatory "SO THERE! "

Edit 2: In seriousness, what I actually had envisioned is probably just a glorified A* once I can translate potential actions into some approximate completion time (to compare the efficiency of grinding in location x to completing quest y).. which may or may not qualify as a planner.

I believe the word you're looking for is 'arrogantly'.

[namreeb]

Hmm, I don't think so! I know we had discussed that topic before!

[namreeb]

Necro-bump. I just realized Alex Champandard found my spam trap email address somehow and sent me a message asking me to add the quote I just added in my initial reply to this thread.

[Cypher]

Alex is watching! Quick, everybody look busy.