From execution traces to specialized inference
Abstract/Contents
- Abstract
- We often envision a future where computers answer our questions. This is inference: coming to conclusions about the world based on a limited amount of information. One formulates a hypothetical world (model), then simulates and analyzes its behavior with respect to evidence. Probabilistic programming languages are a recent approach where we can express any hypothetical world as a program with random choice primitives. However, this descriptive power often sacrifices performance of inference. In my work, I explore execution traces as a solution that enables high performance while preserving descriptive power. This leads to three subsequent developments. The first is Shred, a tracing compiler that generates efficient Metropolis-Hastings MCMC code from probabilistic programs. Performance was seen to be competitive with hand-coded MCMC in some cases. Like the tracing just-in-time (JIT) compilers that served as the inspiration, there is a sacrifice of efficiency in representing multiple control flow paths. Unlike with traditional JIT compilers, the execution target of probabilistic programming language traces is not limited to straightforward execution on a low-level language, but also includes state-of-the-art inference engines. I developed Solitaire, a language for procedural content generation with constraints, combining trace graph compilation with SMT solving and probabilistic languages. Finally, rather than being limited to an intermediate representation, execution traces can also be treated as abstract objects of inference. The third development is model accretion, a stochastic search-based MAP inference algorithm that improves performance of procedural content generation by re-using previous executions.
Description
| Type of resource | text |
|---|---|
| Form | electronic; electronic resource; remote |
| Extent | 1 online resource. |
| Publication date | 2015 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Associated with | Yang, Lingfeng |
|---|---|
| Associated with | Stanford University, Department of Computer Science. |
| Primary advisor | Hanrahan, P. M. (Patrick Matthew) |
| Thesis advisor | Hanrahan, P. M. (Patrick Matthew) |
| Thesis advisor | Aiken, Alexander |
| Thesis advisor | Goodman, Noah |
| Advisor | Aiken, Alexander |
| Advisor | Goodman, Noah |
Subjects
| Genre | Theses |
|---|
Bibliographic information
| Statement of responsibility | Lingfeng Yang. |
|---|---|
| Note | Submitted to the Department of Computer Science. |
| Thesis | Thesis (Ph.D.)--Stanford University, 2015. |
| Location | electronic resource |
Access conditions
- Copyright
- © 2015 by Lingfeng Yang
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...