<?xml version="1.0" standalone="yes"?>
<Paper uid="P04-3032">
  <Title>Dyna: A Declarative Language for Implementing Dynamic Programs[?]</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> Dyna is a declarative programming language for building efficient systems quickly. As a language, it is inspired by previous work in deductive parsing, adding weights in a particularly general way. Dyna's compiler has been designed with an eye toward low-level issues (indexing, structure-sharing, garbage collection, etc.) so that the cost of this abstraction is minimized.</Paragraph>
    <Paragraph position="1"> The goal of Dyna is to facilitate experimentation: a new model or algorithm automatically gets a new memory layout, indexing, and training code. We hope this will lower the barrier to entry in the field, in both research and education. In Dyna we seek to exploit as many algorithmic tricks as we can, generalizing them to as many problems as possible on behalf of future Dyna programs.</Paragraph>
    <Paragraph position="2"> In turn the body of old programs can provide a unified testbed for new training and decoding techniques.</Paragraph>
    <Paragraph position="3"> Our broader vision is to unify a problem's possible algorithms by automatically deriving all of them and their possible training procedures from a single high-level Dyna program, using source-to-source program transformations and compiler directives. We plan to choose automatically among these variants by machine learning over runs on typical data. This involves, for example, automatically learning a figure of merit to guide decoding.</Paragraph>
    <Paragraph position="4"> The Dyna compiler, documentation, and examples can be found at www.dyna.org. The compiler is available under an open-source license. The commented C++ code that it generates is free to modify.</Paragraph>
  </Section>
class="xml-element"></Paper>