Soundbites

• The RALPH architecture was motivated by abstract concerns of bounded optimality, specifically: the need for rational control of reasoning when reasoning is an expensive act;....
• The Multiple EAs framework helps to alleviate the knowledge acquisition "bottleneck" with its multiplicity of knowledge types.
• The principal job of the agent is to make the best decisions it can as soon as it can, given that making optimal decisions is usually out of the question....This requires two things: first, that the agent have some high-value computation steps available within the architecture; second, that it be able to select them.
• Perhaps not surprisingly, more empirically motivated research in the SOAR project has generated many of the same design conclusions, including explicit metareasoning about base-level computations and immanent compilation.
• Differences include the lack of the full range of EAs in these architectures, the lack of the appropriate decision-theoretic connection between metalevel and object level computations, the lack of an immanent inductive process, and the fact that the RALPH architecture hasn't been built yet.