Embodied Agents
By ai-depot | June 30, 2002
Challenges & Applications
Challenges
Despite the many advantages that such embodied agents provide, there are some crucial issues that are worth mentioning.
Efficiency
You might think that simulating such a system has a computationally expensive overhead. Indeed, this is true to some extent — there’s no point lying about it. However, as for any other computer science problem, there are many ways of optimising it to shreds, making sure to minimise the computation. This includes techniques such as:
- Lazy Evaluation
- Event-Driven Processing
- Function Inlining
Also, designing the AI around the limitations can get the final speed close to what it normally is. First, trying to minimise the calls to expensive procedures, which is good practice anyway. Second, spreading the computation over numerous frames, which is also common policy (as well as providing more realistic behaviours). Third, working on a level-of-detail basis, where only the necessary information is queried.
Authenticity
When there’s a thorough specification of what data in the environment you can access, and a reasonably restricted set of actions to perform, designing the AI can be an “interesting issue” ;) Indeed, hackers or game developers are extremely used to exploiting short cuts — and justifiably so. Despite this approach being based on genuineness, it can handle dishonest models too.
Additionally, there are many benefits to the embodied approach. Not only does the robust specification make the development much simpler, but it also implies a much easier way of obtaining believable behaviours.
Realism
Defining a biologically inspired interface is only half of the problem. Scripted behaviours would provide satisfactory results, though not in all cases. Agents really need plausible underlying algorithms as well, in order to complete the illusion. This involves using human-like heuristics, and also potentially machine learning algorithms.
Applications
Computer Games
You’ve most likely noticed the constant references to the computer games industry, and obviously this technology applies will to them. Realism and modularity are two of the main benefits, and as more power can be sacrificed to the AI, accurate simulation of virtual bodies will become a necessity - and possibly even a standard!
Behavioural Research
Due to the realism of the solutions, much academic research is based on genuine animats. By actively enforcing constraints on the agent and even providing biologically accurate perception errors, this makes the study of behaviours even more accurate.
Robotics Experiments
When the animats are fully constrained with respect to their virtual world, and if this environment is an accurate model of real-life, techniques applied with generally transfer to physical robots. It’s a subject of debate among robotics researchers, but general consensus is that “Simulators are doomed to succeed.” — in the words of Rodney Brooks. Realism is increasing drastically along with computational power, so this is looking more and more likely.
Conclusion
Generally speaking, embodied agent will fast become the orthodoxy for simulated animats. The benefits greatly outweigh the disadvantages, and once we have developed underlying algorithms of satisfactory quality as well as a framework to simplify development, there will be no obstacles in our way.
The question is, does the short-term future of game AI involve embedded agents? Since answering anything about the future is purely based on faith and beliefs, I can only give you my personal thoughts and observations…
I believe embedded agents genuinely are a valid option for next-gen, or even games currently in development. In the immediate future, until we get the technology polished, it is possible to change our policy of game AI; we can develop agents that behave in a naturally stupid fashion — not being incredibly smart in some cases, but acting realistically in all situations. This doesn’t necessarily involve online learning, but a more robust AI framework.
In the medium term, I think and hope there is a miracle solution around the corner. A learning solution is most definitely probable, which would be able to deal with embodiment intrinsically, and provide behaviours rivalling the “rigid” dishonest game AI in their areas of predilection, and surpassing them in general situations.
To wrap up, here’s a quick plug of my latest project: an open source framework which aims (among other things) to embody game animats. This embodies all my hopes in this paradigm — pardon the pun. Hopefully, this will go a long way to prove to developers that this is a viable solution (a guinea pig if you will :), but also support the hobby AI development community (making their task of creating realistic agents so much easier).
Written by Alex J. Champandard.
Pages: 1 2
Tags: none
Category: essay |
