The Secret Agent Man

How NeuroStudio Works

I spoke with Walter Alden Tackett, President and co-founder of Neuromedia, Inc., and he told me that "…the bottom line is that NeuroStudio is highly optimized to run fast and be easy to use, while still being able to track the context of a conversation. Much of the [information] necessary to track context is intelligently extracted by the compiler, which is also responsible for making responses require minimal search time - approximately O(log(log(n))) to search the topics for an answer, where n is the number of topics in the bot. As far as we know, it is the only AI system ever designed which is precisely optimized from the ground up for use with chat bots and many simultaneous conversations."

Why is NeuroStudio so fast and how can it track the context of a conversation? Well, I asked Ray Dillinger, who goes by the title of "Botmaster," at Neuromedia. According to Ray, "The way a typical chat bot engine works is to check user input against ‘patterns’…[which] are associated with sets of "actions"; usually the definitive action is saying something in response…A ‘topic’ is what we call the association between a pattern and a set of actions."

If a chat bot engine attempts to match inputs with topics in a sequential (fixed) order, the bot will be unable to answer questions based on the context of the question. It will simply respond to a query with the first matching action it encounters. Secondly, as the number of topics in a knowledge base grows, finding associations between a pattern and action will require ever increasing amounts of amounts of time and processing power.

To solve these problems, NeuroStudio uses a combination of patent pending techniques. Ray first pointed out that with NeuroStudio "…topics are considered in an order that depends on the course of recent conversation… When a particular topic is selected and its actions performed, its ‘subjects’ are activated. And on the next input, the topics are checked in order from most recently activated subject to least recently activated subject. Also, we have one special matching operator, ‘focused’, …for tightly context-dependent [material]."

NeuroStudio also automatically calculates a "specificity" score for each matching pattern to handle situations where some questions are more specific than others. This lets a bot developer write extremely broad and general responses to questions, which can be used when nothing more specific can be found. The subjects mechanism works in conjunction with specificity score to enable context-dependent but also less specific conversation between a user and a bot.

To conquer the processing time and power problem for large topic files, NeuroStudio uses a highly innovative approach. As Ray puts it, "The topics are sorted into a hash-table keyed by the simple matching patterns. In short, the system starts by finding out what simple patterns match, and uses them as keys into the hash table to find more complex patterns built of them, and uses the set of those patterns that match as keys to find more complex patterns…" until a topic is matched.

Ray goes on to say that as a result of the hash-table approach, NeuroStudio "…never [has] to repeat computations…This is why a NeuroStudio bot with 1500 topics (or 15,000 if you've got the memory space) can respond *almost* as fast as a NeuroStudio bot with 300 topics."

To verify this claim, I contacted a NeuroStudio user. William A. Orlowski is the developer of Brain (http://orlo.emi.net/html/brainframe.htm), a chat bot that draws on more than 10,000 topics for its ability to answer questions on an amazingly wide range of subjects. He said his bot gets more than 200 visitors a day and up to 25 concurrent sessions at a time. William also told me that this NeuroStudio bot runs on a PC with an Intel 180 MHz MMX Overdrive processor, 40 MB RAM, 2.6 GB hard drive, and Windows 98, over a 56KB modem connection to the Internet. Despite the lack of a sizeable pipe to the Net or a server operating system, Brain was quick to respond and even provided me with images rapidly, as illustrated in Exhibit 6.

Return