The Two Approaches
I have been working in a direction that is relatively risk. It is risky because it is unclear whether the direction will work and whether the direction has benefits.
The literature on artificial intelligence has suggested that the direction is worth taking. Some of the areas that the project has researched include natural language processing, commonsense knowledge bases, cognitive architectures, question answering systems, document summarisation systems, embodied artificial intelligence and artificial life.
The direction is as follows. Researchers can either program the exact behaviour they want or take a developmental approach.
It sounds simple. The decision between each approach depends on one's purpose.
It helps to have a specific way to analyse intelligence. The area of natural language communication is a good test for capability of each approach. This is because it reveals what each approach is capable of doing.
I have been writing a discussion on the differences between the programming approach and the developmental approach because it has been a challenge to communicate the differences between the approaches.
"What's next?"
Right now, I am asking, "What's next?" I want to move on to the more practical stage of my research. I have read papers where researchers created agents that developed their own language.
This is important because of two things. Firstly, it shows that with an appropriate cognitive architecture and an appropriate environment, agents can develop communication. This is important because it is a potential way to overcome the problems of programming desired abilities. Secondly, once agents can communicate, they can gain from indirect experience. This is important because it allows them to build up knowledge. This is different to just having stored information. For example, there is a difference between having an engineering textbook in your hand compared to having studied the textbook and being able to use the principles.
Back to my question, "What's next?" The research papers indicate that the next challenge is to scale up. Now I have two worries. Firstly, I am worried that my ideas are not new. A requirement of a PhD topic is that it is new. I do think that research that scales up current research is new even if it is only a small step. Secondly, I am worried about how to do the scaling up. In the past, I usually relied on thinking that I am intelligent and I tried to innovate. This time I am instead going to do the wise thing of first imitating.
My first idea was to spend weeks or even months scaling up the research to a level that I thought would be useful before producing results. I wanted to create a sophisticated simulated embodied agent to interact with a complex environment. I wanted the agents to have splat vision sensors, a sophisticated body capable of human like movement, sensitive skin, and human like vocal systems, hearing and so on. I was considering using a game engine and the complexity grew and grew. This would have entailed a lot of work that may or may not have paid off.
Instead, this time, I am going to take a close look at what other researchers are doing. One of the papers that I have read mentioned a paper that surveyed toy world systems and another paper that investigated not only emergence of a simple vocabulary and one word communication but also the emergence of multi-word communication.
In addition, there were certain points in the papers that I felt uncomfortable with and that contrasted with certain ideas that had formed from my prior readings.
What environments are available for this type of research?
There are not many specific environments. There are environments for artificial life research but that is about it.
The environments that relate to embodied AI mainly involve developing control systems to guide movement.
Novamente and the related projects at the Artificial General Intelligence Research Institute are interesting projects and Ben Goertzel is one of my heroes. However, they conflict with some ideas that I have about evolved and emergent intelligence and communication. In addition, Novamente is a commercial product.
The Reinforcement Learning Toolbox is an interesting project. By itself, it is more of a disembodied approach and not interesting but it looks like it can connect with simulators for an embodied approach.
Two of the interesting simulators the Reinforcement Learning Toolbox connects with are Webots and Hoap2. Webots simulates physics in 3D and Hoap2 is a simulator for a humanoid robot done in Webots.
Thyrix is a commercial simulator that seems to be a bit simple.
Tim Taylor is a name that I have seen mentioned in a number of places.
"Evolving Virtual Creatures in 3D Simulations," describes a number of artificial life projects that relate to embodied artificial intelligence. The page is also available at AI Game Development. These projects focus on low-level behaviour such as movement rather than high-level behaviour such as communication and knowledge building.
One of the links is J. Bongard's PhD thesis, "Incremental Approaches to the Combined Evolution of a Robot's Body and Brain." Bongard co-authored the book great book, "How the body shapes the way we think."
Another one of the links from AI Game Development is to Russel Smith's PhD thesis, "Intelligent Motion Control with an Artificial Cerebellum." Again, I wish it touched on evolutionary emergent intelligence and communication in embodied social agents.
The iCub project on the other hand goes in the direction of actual robot building. I am not a big fan of robot building because I admit I do not have the skill and I like to tweak at the speed of thought.
Framsticks is another simulator that seems to focus more on the artificial life aspect rather than evolutionary emergent intelligence and communication.
Gene Ruebsamen has an interesting page that focuses again on the artificial life direction.
The Flexible Embodied Agent aRchitecture seems to be a disembodied approach. It has a cool acronym but details on the project seem a bit sketchy on the SourceForge page. The actual website for the project is jam packed with information.
There are some materials on "simulated robotics."
Microsoft Robotics Studio - Official Page
One of the things I am looking for is an architecture that is similar to multi-player games. The idea is for the server to take care of simulating the environment and each client machine to take care of processing the artificial intelligence.
One of the ideas that occurred to me was to have the experiment in an environment such as Second Life. Researchers and people who are just having fun have also given this idea of artificial life in Second Life a go.
I prefer a humanoid morphology. I also prefer a more natural fitness function. I want the a sophisticated human like environment to provide an appropriate fitness function rather than the abstract ones commonly used such as getting to a location.
Although the experiments were quite impressive, they were also quite simple.
What do the research papers suggest?
It kind of sucks but it looks like most of my research today has not been that useful. I should know by now that the there is not that much information on evolved emergent intelligence and communication in embodied social agents.
"The Adaptive Advantage of Symbolic Theft Over Sensorimotor Toil: Grounding Language in Perceptual Categories" by A. Cangelosi and S. Harnad, 2002, was a great paper. The paper is available at cogprints and at citeseer. Cangelosi has had a number of publications in this area and papers often quote Harnad in regards to the symbol grounding problem.
In regards to ideas on how to scale up it is not so clear. Cangelosi and Harnad did refer to a number of key papers such as one on compositional languages, Cangelosi, 2001.
However, I had hoped one of the papers they referred to, Knight et al., 2000, was in regards to a survey of toy worlds but it turned out to be a text discussing evolutionary languages. It may be useful but seems too focused on theory and not from a computer scientist's point of view.
The book is available at Amazon and at Google Books. I checked in my library and found some other interesting related books too. The title is "The Evolutionary Emergence of Language: Social Function and the Origins of Linguistic Form."
Evolutionary phonology: the emergence of sound patterns by Juliette Blevins, 2004
Language acquisition, change and emergence: essays in evolutionary linguistics edited by James W, 2005
Why we talk: the evolutionary origins of language by Jean-Louis Dessalles; translated by James Grie, 2007
Biology, evolution, and human nature by Timothy H. Goldsmith and William F. Zimmerman, 2001
Cognition and material culture: the archaeology of symbolic storage edited by Colin Renfrew and C, 1998
The emergence of agriculture: a global view edited by Tim Denham and Peter White, 2007
The seeds of speech: language origin and evolution by Jean Aitchison, 1996
I found these books with the keywords, "The evolutionary emergence of."
On closer look, Knight et al., 2000 looks like are really good book. It presents a number of simple abstract games to demonstrate some important ideas. There are also a number of exciting theories regarding ideas such as politics and language. Yeah baby!
Okay, to stop myself from letting the excitement carry me away, these do not directly suggest any ideas on what to do next in regards to my research.
One of the things that has got me worried
I notice that a large number of the research papers focus on one concept that researchers have polished to a razor's edge. For example, researchers have focus on locomotion or categorisation or basic communication. Rarely have the papers discussed a scaled up experiment.
However, I notice that the disembodied research approach does discuss quite scaled up projects in the area of question answering.
I like the title "Evolutionary Emergence of Language in Simulated Embodied Social Agents."
No comments:
Post a Comment