Management of speedup mechanisms in learning architectures.

详细信息   

• 作者：Cheng ; John Christopher.
• 学历：Doctor
• 年：1995
• 导师：Mitchell, Tom
• 毕业院校：Carnegie Mellon University
• 专业：Artificial Intelligence.;Computer Science.
• CBH：9522944
• Country：USA
• 语种：English
• FileSize：6980145
• Pages：170

文摘

Learning architectures typically operate rather inefficiently. To increase performance, two strategies are commonly used: speedup mechanisms are incorporated into the architecture, and architecture operation is simplified. Unfortunately, both these strategies have drawbacks.;Because of the utility problem, inappropriate use of speedup mechanisms can actually decrease system efficiency. Hence, good speedup mechanism management--deciding when, where, and which speedup mechanism to use is--important if the mechanisms are to be effective. Typically, however, good management strategies are not available. Architecture-provided strategies are usually very simple. and cannot use the mechanisms appropriately all the time. Good user-provided strategies are also difficult to develop--under a complex system or domain, it can be difficult to understand system behavior well enough to specify a good management strategy. Furthermore, user or architecture-provided management techniques are usually fixed, and cannot adapt to environment dynamics. Hence, lack of good management strategies limit the effectiveness of speedup mechanisms.;Simplifying an architecture's inference mechanism yields dramatic efficiency gains. Unfortunately, gaining efficiency in this manner usually sacrifices fine-grain control over the behavior of the system, or architecture flexibility. Consequently, this speedup technique forces the domain designer to operate at the flexibility/efficiency tradeoff point chosen by the architecture. This thesis investigates ways of handling both of these problems. The speedup mechanism management problem is approached by making the architecture itself responsible for developing a management strategy. An agent, embedded into the architecture, observes system operation, invoking speedup mechanisms appropriately. This approach allows the architecture to tailor its strategies individually to different domains, increasing speedup mechanism usefulness. Furthermore, because the agent can monitor the architecture continuously, it can adapt its management strategies to the dynamics of the environment.;This dissertation also presents an algorithm that can be used to reduce the flexibility/efficiency constraints on the domain designer, giving him more options. The designer is allowed architecture flexibility, but if flexibility is not needed, the unnecessary flexibility is automatically traded for efficiency.