direct optimization
play

Direct Optimization CSC2547 Adamo Young, Dami Choi, Sepehr Abbasi - PowerPoint PPT Presentation

Direct Optimization CSC2547 Adamo Young, Dami Choi, Sepehr Abbasi Zadeh Direct Optimization A way to obtain gradient estimates that directly optimizes a non-differentiable objective. It has first appeared in structured prediction


  1. Gumbel Process

  2. Gumbel Process We know:

  3. Gumbel Process We know: Therefore:

  4. Gumbel Process We know:

  5. Gumbel Process A B

  6. Gumbel Process A B

  7. Gumbel Process A B

  8. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value.

  9. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value.

  10. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value.

  11. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value. 1.3

  12. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value. 1.3

  13. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value. 1.3

  14. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value. 1.3 1.3

  15. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value. 1.3 1.3

  16. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value. 1.3 1.3 1.1

  17. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value. 1.3 1.3 1.1

  18. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value. 1.3 1.3 1.1

  19. Trajectory Generation ● Lazily create partitions of trajectories. ● Recursion rule: ○ For , copy parent node’s value. ○ For the remaining choices of actions, group them and compute truncated value. 1.3 1.3 1.1 1.3

Recommend


More recommend