|Monday, July 12, 2004|
|Ai releases HAL2004|
|Savyon, Israel -- July 12, 2004 - Artificial Intelligence Research (Ai) |
(http://www.a-i.com) announced today the release of a new version of its Learning Algorithm, HAL2004.
HAL2004 has been implemented in the Java programming language. Previous engines (HAL and HAL2003) were implemented in TCL (a scripting language). The switch to a compiled system provides significant performance improvements: HAL2004 is 10-50 times faster than HAL2003.
Previous HAL versions had severe difficulties handling personalities with tens of thousands of turns. HAL2004 now loads the largest available personality in under 10 seconds and maintains an average turn response time of well under 2 seconds.
The runtime memory consumption of HAL2004 has also been greatly improved. While HAL2003 needed about 1.2GB of virtual memory to run the largest personalities, HAL2004 uses less than 250MB. This reduction in memory requirement allows for more personalities to be trained concurrently.
New State File Format
HAL2004 uses standard XML format to store the personality state between sessions. Previous versions used a text file formatted in a TCL internal format.
The switch to XML has forced the definition of a schema for the state. The schema has been defined in terms of the knowledge acquired by the engine rather then the format it uses to navigate in. This allows for feeding the saved state into newly developed algorithms in the future.
By using XML, the issue of international character support has also been addressed. The XML is saved using UTF-8 encoding and can therefore accommodate the full UNICODE character set. This attribute, coupled with the international character set support of the Java program language, results in a fully character set independent engine.
Enhanced Learning Algorithm
HAL2004 provides better dialog performance - fewer corrections required to learn new behavior, and includes new classification and generalization capabilities.
Regression Dialog Suite
HAL2004 is continuously tested using a jUnit test suite, which includes tests for individual implementation classes. The suite includes a set of test dialogs with respective expected behavior. With this new facility, changes in the algorithm are always verified against a known set of dialogs. This provides for greater release stability and consistency along with a mechanised way to regenerate field problems and verify their resolution.
HAL2004 is now open to the public. To create a new HAL2004 personality, ask Alan to speak with a fresh copy of HAL.
Artificial Intelligence Research (Ai) is a groundbreaking research project, focusing on creating genuine Artificial Intelligence – the technology that enables machines to converse with humans in natural language. Ai's research facility is located in Savyon, Israel.