SGER: An Architecture of Diversity for Common Sense Reasoning

人工智能研究在计算理解人类推理方面取得的进展有限：我们未能强调在面对不熟悉的问题时创建具有“智谋”的程序。在计算理论和应用方面，重点主要放在特定的算法和数学技术上；复杂系统很少产生可重复的结果或对系统构建理论的基本见解。资源丰富的程序将需要应用多种推理技术，一种用于协调活动的分层体系结构，通过这种结构，高阶反思过程可以诊断低阶审议过程中的故障。需要“常识性推理”的场景提供了一个足够范围的铁砧，只有资源丰富的程序才能满足需求。两个项目支持这些目标：Roboverse建立了一个虚拟世界，人形角色参与复杂的多领域场景，将常识应用于身体、社会和心理活动；类比是多种表示之间相互联系的规范，结合了类比、案例推理、统计分析和逻辑推理等多种推理技术。Panalogy的早期版本实例化了一组元管理组件，演示了协调和修复基线推理过程的不同方法。这些资源通过为实验再现和比较提供新的理论原理和评估工具和方法的激励演示，使更广泛的研究界参与。我们的目标是通过研讨会、出版物和直接参与，鼓励企业和政府提供更广泛的支持。这种在资源计算研究方面不断增加的努力，为那些与人类一起寻找对科学和社会重要的复杂问题的解决方案的程序开辟了道路。

Artificial Intelligence research has made limited progress towards computational understanding of human reasoning: we have failed to emphasize the creation of programs with "resourcefulness" in the face of unfamiliar problems. In computing theory and applications the emphasis has largely been on specific algorithms and mathematical techniques; complex systems have rarely yielded reproducible results or fundamental insights into a theory of system construction. Resourceful programs will require the application of multiple reasoning techniques, a layered architecture for coordinating activity whereby higher-order, reflective processes can diagnose failures in lower-order, deliberative processes.Scenarios requiring "commonsense reasoning" provide an anvil of sufficient scope that only highly-resourceful programs will meet the requirements. Two projects support these aims: Roboverse establishes a virtual world where humanoid characters engage in complex, multi-realm scenarios, applying common sense to physical, social and mental activities; Panalogy is a specification for interconnect among multiple representations, combining diverse reasoning techniques such as analogy, case-based reasoning, statistical analysis and logical inference. An early version of Panalogy instantiates a set of meta-management components, demonstrating different ways to coordinate and repair baseline reasoning processes.These resources enable wider research-community involvement by providing motivating demonstrations of new theoretical principles and evaluation tools and methodologies for experimental reproduction and comparison. We aim, through workshops, publications and direct engagement to encourage broader support from corporate and governmental sources. This increased effort in resourceful computing research leads the way towards programs that engage readily with humans in the search for solutions to complex problems of importance to science and society.