71 replies to this topic

[IADaveMark]

Holy autonomous agents, Ferretman! Have you looked at the schedule for this year? This is not bad at all! (Other than the fact that they misprinted one of your sessions again and scheduled 2 on Wednesday.) Not only do we have your 3 roundtables... There are 3 days worth of serial sessions on "Embodied Autonomous Agents", Brian Reynolds has a session on "AI and Design: How AI enables designers", There''s one entitled, "Would the real emergent gameplay please stand up", There''s one on Friday called "Beyond Finite State Machines: Managing Complex, Intermixing Behavior Hierarchies" That''s 8 different sessions and, assuming they fix your schedule, they are all at different times! Not too freakin'' bad! Dave Mark - President and Lead Designer Intrinsic Algorithm - "Reducing the world to mathematical equations!"

[alexjc]

There''s also Bryan Stout''s lecture on potential fields for reactive navigation I believe.


Hang on, deja vu

That''s what you get for hacking a BB by posting code to it!!

[Timkin]

Would someone mind telling me what happened to the posts in this thread please?

[IADaveMark]

Odd... it seemed that once the thread got completely off topic and rediculously tainted, everything just evaporated. Certainly these isn''t anything valuable missing, is there?

Dave Mark - President and Lead Designer
Intrinsic Algorithm - "Reducing the world to mathematical equations!"

[Timkin]

Just because a thread gets off topic isn''t sufficient cause to delete certain posts in it. The ONLY reason posts get deleted is if they are in serious violation of the GD.net TOS. Usually in that instance, the thread either gets closed or it gets deleted and moved to a special forum viewable by moderators only so it can be reviewed. If it''s just a single post, then a moderator can either edit the offensive portion out or delete the post. In this instance, every single reply was deleted. If a moderator has deleted these posts from within the thread, simply because they felt they were off topic, then that person has seriously breached the boundaries of their responsibilities. If it wasn''t a moderator, but rather a forum bug, then we need to identify it. If it was neither, then it was likely a forum hacker. Again, we need to ascertain whether this is the case.

Was the content important? That''s not really for one person to decide. I certainly don''t like having my posts deleted for no reason. I''m sure others feel the same way. I''d like to get to the bottom of this, since in Ferrettman''s usual absence, I take responsibility for these forums (actually, Ferrettman hardly ever shows himself in these forums these days...)

Timkin

[alexjc]

heh, well the moderators had the choice of dodgy statements to justify their actions. My moody initial reply got everything off on the wrong foot . Appologies!

Alex

[Krippy2k]

It''s that new autonomous forum moderator Either that or my friend''s old boss is in the homeland security h4x0r department now and didnt like me revealing his under-the-fuselage antics.

Come to think of it, he kinda looked like Tom Ridge, in a neckless kinda way.

Peace

[IADaveMark]

Hey Alex... does this sound familiar?

quote:

---

Session Title: Embodied Autonomous Agents

Description: Autonomous agents are entities that perceive their world and act within it. Embodied AAs have their world equal to the real world, in which they move, sense and are sensed, act and are acted upon. In our virtual game worlds, they are the entities we see on the screen, acting within the small worlds visualized in the display, as opposed to agents that operate on other levels, such as making strategic decisions or giving advice to players.
The demands on embodied AAs are high: they are now expected to act only on information they should reasonably perceive; their action animations have ever-growing standards of realism; they may even be expected to show different attributes and personalities from each other. This roundtable provides the opportunity to discuss the issues developers face when implementing embodied AAs, and share perspectives and possible solutions.
Each day of the roundtable will focus on a different stage of the "perceive-think-act" cycle. The first day focuses on sensing and perception, and issues such as data transmission (line-of-sight checks alone can take up more than the rest of the AI combined), data content (what exactly is sensed), queries vs. interrupts, and focus of attention. The second day focuses on decision-making, and issues such as agent memory, individual world representations, beliefs, goals, personality, etc. as well as reasoning techniques. The final day covers action, and issues such as action representation, control over the animation, realistic action, physics, collision detection, etc. The specific issues discussed each day depends on the desires and contributions of the participants.

Takeaway: The attendees come away with a better appreciation for the issues involved with embodied AAs: the problems developers have to deal with in their implementation as well as the solutions and ideas to try that they share with each other.

Intended Audience and Prerequisites: Programmers, animators, designers, or any other developers who wrestle with the issues involved with implementing embodied AAs.

---

Dave Mark - President and Lead Designer
Intrinsic Algorithm - "Reducing the world to mathematical equations!"

[IADaveMark]

quote:

---

Original post by alexjc
There''s also Bryan Stout''s lecture on potential fields for reactive navigation I believe.

---

Alex, did you listen to that audio from Eric''s roundtable session last year? Isn''t this along the lines of what we were talking about with the soccer/football player navigating toward the goal with the ball?



Dave Mark - President and Lead Designer
Intrinsic Algorithm - "Reducing the world to mathematical equations!"

[IADaveMark]

quote:

---

Original post by Timkin
Would someone mind telling me what happened to the posts in this thread please?

---

To quote from a famous book by a famous author:

"Contradictions do not exist. If you think you have found a contradiction, check your premises and you will find one of them to be in error."

Check your premises, Timkin. Nothing happened to any posts in this thread. Move on.

Dave Mark - President and Lead Designer
Intrinsic Algorithm - "Reducing the world to mathematical equations!"

[IADaveMark]

I rather like the concept in this one as it is my core belief and founding principle of Intrinsic Algorithm.

quote:

---

AI and Design: How AI Enables Designers
Speaker(s): Brian Reynolds

Description: Whether you’re a programmer who’s always wanted to work on the game design or a designer who thinks there might be something to this “programming” thing, here’s your chance to talk with someone who has worked both sides of the fence. We’ll focus on AI and the ways in which AI development does (or should) overlap with the game design process, drawing case studies from the presenter’s experiences as Lead Designer for Rise of Nations, Alpha Centauri, and Civilization II. We’ll talk about why delaying AI development “until the design docs are final” is a wasted opportunity, and how both AI and Design benefit from simultaneous prototyping. We’ll explore not only the traditional use of AI to determine goals and strategy for computer players, but also the critical role of AI in supplying “personality” to computer-controlled characters. Perhaps most importantly we’ll talk about the sometimes-unexpected ways AI techniques can be invaluable in content generation.

Idea Takeaway: New ways to integrate game design with AI development, and how AI can be uniquely valuable in content generation.

Intended Audience and Prerequisites: Programmers, Designers, and especially Programmer-Designers! Knowledge of basic C++ programming techniques is helpful but not essential.

---

Dave Mark - President and Lead Designer
Intrinsic Algorithm - "Reducing the world to mathematical equations!"

[alexjc]

quote:

---

Original post by InnocuousFox
Hey Alex... does this sound familiar?

---

Indeed, it''s one of my core beliefs and also one of the premises of my book!

I have one major concern with the format; the sense-plan-act model - as a legacy of the 1970s - is particularly dated. I think a distributed AI that reacts to specific stimuli with small event handlers is much more suited to games (in terms of efficiency and simplicity). So instead of having one monolithic architecture, you have a distributed hierarchy of reactive components... sound familiar too?

Maybe someone can bring it up?

Alex

[IADaveMark]

quote:

---

Original post by alexjc
Maybe someone can bring it up?

---

Not me, dude! Anyway, it''s your buddy Mr. Stout... YOU bring it up!



Dave Mark - President and Lead Designer
Intrinsic Algorithm - "Reducing the world to mathematical equations!"

[alexjc]

quote:

---

Original post by InnocuousFox
Not me, dude! Anyway, it''s your buddy Mr. Stout... YOU bring it up!

---

I would if I was going! It makes sense to keep the 3 roundtables, but it''d be worth discussing...

Alex

[Timkin]

quote:

---

Original post by alexjc
I have one major concern with the format; the sense-plan-act model - as a legacy of the 1970s - is particularly dated.

---

I disagree that this model is ''dated''... it has merely evolved into:

                                ---> act --
                              /             \
--> sense -> revise beliefs ==               ----------------->
                              \                            /
                                ---> evaluate --> replan --
 

where most architectures interleave metadeliberation with acting. Architectures like the one I designed (PPR) allow for seemless replanning while concurrently revising beliefs, deliberating and acting, along the lines of:

      --------------------------
     |  act                     |
     |                          |
---> |  sense -> revise beliefs | --->
     |                          |
     |  evaluate -> replan      |
      --------------------------
 

So, my point is, that distributed architectures of heirarchical subsumptive elements are just one model for situated agents. Traditional approaches have evolved a lot since the 1970s and are still very applicable techniques... even more so where one requires guarantees about the global optimality of actions and plans; something that reactive architectures cannot provide and never will (by definition).

Cheers,

Timkin

(Dave: On the point about ''letting go''... I have a job to do here that extends beyond any desire to find out what happened to my posts... if someone is hacking the forums, I need to know so I can inform the GD.net staff. If it''s a rogue, or naive mod they also need to know about that.)

[IADaveMark]

No one is hacking the forums. At least not in this case.

Dave Mark - President and Lead Designer
Intrinsic Algorithm - "Reducing the world to mathematical equations!"

[alexjc]

Timkin, I guess we have different backgrounds, so we''re unlikely to agree -- but here goes.

quote:

---

Original post by Timkin
I disagree that this model is ''dated''... it has merely evolved

---

That''s a different model, it was invented because the original is so dated

quote:

---

So, my point is, that distributed architectures of heirarchical subsumptive elements are just one model for situated agents.

---

I just to be clear, I never said or implied subsumption. It''s quite limited... and dated too! 1980s this time, hehe.

quote:

---

Traditional approaches have evolved a lot since the 1970s and are still very applicable techniques... even more so where one requires guarantees about the global optimality of actions and plans; something that reactive architectures cannot provide and never will (by definition).

---

I have no doubt they are applicable, as your Ph.D. research shows. But what was the context of your research: single autonomous agents controlling flight -- if I remember correctly?


But in games, who "requires guarantees about the global optimality of actions"? My argument is that this is not at all suitable in two contexts relating to games:

1) Realtime systems with highly constrained resources

If you look into robotics research over the past decade, you won''t find many pure deliberative architectures like the ones you described. They are mostly designed as reactive architectures. Granted, some have planning components (definitely useful),but the overall design is usually based on distributed components -- making it easier for the system to react to important stimuli (and easier for the engineer to design the system that way).

The process of encapsulating deliberative techniques within reactive components and then assembling them together has become the defacto standard for producing useful robot AI with limited resources. Games are no different.

2) Synthetic creatures that do not try to solve narrow problems

Having one unique component that solves all the problems just doesn''t scale. Maybe you can get such an agent to control a plane during takeoff, flight and landing, but to get it to behave satisfactorily in most game situations is much harder. You probably could do it with a deliberative architecture, but it''d cost you serious computation!!

I often think of NPC AI architectures as windowing toolkits like Qt. The user and environment are constantly generating new events, which require immediate attention. You assign a message handler to each of these situations, decide if a response is required and delegate the task of producing the answer to the appropriate component in the system (reactive or deliberative). This overall design is a reactive architecture.

The pure deliberative approach would try to take in all the data as it comes in without discarding much, and then run it''s huge process over the whole lot (with optional re-optimization). Appart from being somewhat innefficient, it''s also particularly hard to design and debug (this could be point 3 ).

Anyway, enough from me.

---

AiGameDev.com

[Timkin]

Alex,

quote:

---

Original post by alexjc
I have no doubt they are applicable, as your Ph.D. research shows. But what was the context of your research: single autonomous agents controlling flight -- if I remember correctly?

---

No. The context was decision making and replanning under uncertainty in a dynamic, uncertain environment. Controlling flight was left up to the autopilot... that's what such expert systems are for.

As to your post... let's be clear about a few things... deliberative agents can be formulated from distributed elements. No one should be suggesting that modern deliberative agents are monolithic expert systems. Sense-plan-act in no way implies a monolithic deliberative agent. Indeed, plans of length 1 in the above model are reactive agents. What I was talking about in my post was the distinction between previous models of deliberative agents and current models.

You made the comment that deliberative systems were hard to design and debug (via the assumption that deliberative systems are necessarily monolithic). Reactive systems are far harder to design and debug for a given specification of system requirements. It is certainly not the case in general that you can simply plug modules together and get well-defined, well-behaved agents with known mappings from states to actions.

This is certainly a problem in games with agents built from component systems. The game designer needs to ensure that the agents within the game will behave in a reasonable manner in almost every perceivable context. That's REALLY hard to do unless you either a) severaly limit the contexts available to the agent; and/or b) enforce base behaviours when the agent wants to step outside its permissible state boundaries. A simple example of this sort of problem is utilising an ad hoc fix to agents whos reactive pathfinding gets them stuck in an infinite state loop. This is why we tend to use oracle-like pathfinding algorithms instead of permitting reactive pathing. Such oracles are necessarily deliberative.

Of course I'm not suggesting that deliberative systems are the be-all and end-all of AI or that they have all the answers. Indeed, I believe the answer is necessarily a hybrid of the two, especially for games. We need agents that can formulate plans (that can be perceived and thwarted/aided by players) and act to bring about their goals, but we also need those agents to be reactive to changes in their environment that adversely affect their plans. That was the premise of my PhD research. While you might look at the lower box in my previous post and think of it as a deliberative agent, it's actually both deliberative and reactive. Belief revision gives the PPR agent the ability to perceive changes in the world (and indeed, even in the model governing the world) and this causes the agent to alter its plan if the perception causes the belief that the current plan is of lower value than when it was accepted. The distinction between a deliberative and reactive agent comes in only in the length of the plans considered. Plans of length '1 action' are necessarily reactive behaviours and the system reduces to pure reaction to stimulus.

Timkin

[edited by - Timkin on March 3, 2004 1:52:16 AM]

[alexjc]

I think the issue of design simplicity is worth more discussion, but we can leave that to somewhere else (another thread maybe).


To me, the core issue may be semantics:

quote:

---

Sense-plan-act in no way implies a monolithic deliberative agent.

---

How is that not monolithic? You have one underlying algorithm with tighly integrated data-structures.

And while we're on definitions, do you consider modern deliberative architectures under the Sense-plan-act umbrella?

Alex

[edited by - alexjc on March 3, 2004 8:32:41 AM]

[Timkin]

quote:

---

Original post by alexjc
To me, the core issue may be semantics:

---

No, I think it''s an issue of bias!

quote:

---

Original post by alexjc
How is that not monolithic? You have one underlying algorithm with tighly integrated data-structures.

---

Having one underlying guiding algorithm does not make a system monolithic. Implementation does. As I tried to point out above, deliberative does not preclude distributed implementation... it''s simply that in an agent that performs sense - plan - act in a sequential manner, there is an over-arching control algorithm binding the elements of the implementation together (as there is in most deliberative systems that I have seen...and some so-called reactive systems).

This in no way implies "tightly integrated data-structures". Utilising a central knowledge base and having all components of the deliberative system act on that, or in response to it, is certainly one way to avoid having to share data structures between computational modules. It also provides extensibility and easy improvement to the agent by replacing current modules with new/better methods, as they arise.

quote:

---

Original post by alexjc
And while we''re on definitions, do you consider modern deliberative architectures under the Sense-plan-act umbrella?

---

That really depends on how broad your umbrella is... and who''s work you''re looking at. Certainly, there is still a lot of work being doing in both robotics and soft-bots in sequential, interleaved sensing, planning and acting. However, there is also more work being done every year on parallel deliberative architectures whereby sense, plan and act occur concurrently and usually at different frequencies. So, to answer your question, ''do I consider modern deliberative architectures under the sense-plan-act umbrella'', yes, I do... but I qualify that with the rider that I believe the umbrella encompasses more than the notion of sequential, interleaved sensing, planning and acting and is really an umbrella spanning autonomous agents in the broad sense (and here I mean AAs as defined by communities such as UAI).

For those going to the GDC roundtable... and this is particularly directed at Steve and/or Eric, I''d be interested in hearing the results of discussing this topic:

"Which is better for computer game agents: deliberative architectures or heirarchical, distributed architectures"?

The points to consider should be something like:

Ease of design, implementation, debugging and verification;
Availability and understanding of known methods/algorithms;
Can we simply transplant techniques from robotics into game agents;
What are peoples experiences with either/both paradigms;
What has been successful in the past.


Now, if only there were a way to move part of a thread to another thread!

Cheers,

Timkin