Lecture 8, February 24, 1997

Copyright © 1997 - Brad Myers

Previous Lecture . . . Next Lecture

* Show videos of constraint solvers.*

- Relationships defined once and maintained by the system
- Useful for keeping parts of the graphics together.
- Typically expressed as arithmetic or code relationships among variables.
- Variables are often the properties of objects (left, color)

- Types:
- "Dataflow" constraints; Choices:
- Single-Output vs. Multi-output
- Types: One-way, Multi-way, Simultaneous equations, Incremental, Special purpose
- Cycles: supported or not

- Others: AI systems, scheduling systems, etc.

- "Dataflow" constraints; Choices:

- Simplest form of constraints
`D = F(I1, I2, ... In)`

- Often called
*formulas*since like spreadsheets - ¥ÊCan be other dependencies on D

CurrentSliderVal = mouseX - scrollbar.left

scrollbar.left = window.left + 200

scrollbar.visible = window.has_focus

- Not just for numbers:
`mycolor = x.color`

- Implementations:
- Just re-evaluate all required equations every time a value is
requested
- least storage, least overhead
- Ñ Equations may be re-evaluated many times when not changed.
(e.g,
`scrollbar.left`

when mouse moves) - cycles:

`file_position = F1(scrollbar.Val)`

scrollbar.Val = F2(file_position) - Cannot detect when values change (for redraw)

- Cache current values with each constraint; lazy eval
- Example:
`A = 10`

B = A + 5

C = A * B

D = A + E

E = 20

F = D + C - now need to know when values become invalid and recalculate
in right order

- two phases: invalidate and re-calculate
- invalidate all values that depend on the changed value
- recalculate only values that are demanded
- data structures: depends-on-me, i-depend-on

- may re-evaluate values that haven't changed unnecessarily when conditionals, "max", etc.
- can mark slots/objects that change
- can detect cycles with a counter

- Example:

- Just re-evaluate all required equations every time a value is
requested
- Variations:
- Multiple outputs
`(D1, D2, ... Dm) = F(I1, I2, ... In)`

- Side-effects in the formulas
- useful for creating objects
- when happen?
- what if create new objects with new constraints
- cycles cannot be detected

- Variables in the dependencies:
`D = p^.left + 10`

- important innovation in Garnet we invented, now ubiquitous
- supports feedback objects
- supports loops:
`D = Max(components^)`

- requires the dependencies be dynamically determined

- Constant formula elimination
- To decrease the size used by constraints

- Multiple outputs

- From ThingLab
- Constraints are expressions with multiple variables
- Any may be modified to get the right values
- Example:
`A.right = A.left + A.width - 1`

- Often requires programmer to provide methods for solving the constraint
in each direction:
`A.left = A.right - A.width + 1`

A.width = A.right - A.left + 1 - Useful if mouse expressed as a constraint
- Requires a
*planning*step to decide which way to solve- Many systems compute plans and save them around since usually change same variable repeatedly

- In general, have a graph of dependencies, find a path through the graph
- How control which direction is solved?
`CurrentSliderVal = mouseX - scrollbar.left`

- "Constraint hierarchies" = priorities
- constants, interaction use "stay" constraints with high priority

- Dynamically add and remove constraints

- "Constraint hierarchies" = priorities
- Brad Vander Zanden's "QuickPlan" solver
- Handles multi-output, multi-way cyclic constraints in O(n2) time instead of exponential like previous algorithms

- Required for parallel, perpendicular lines; tangency, etc.
- Also for aggregate's size
- Numerical (relaxation) or symbolic techniques

- Michael Gleicher's PhD thesis
- Only express forward computations
- Tries to get reverse by incrementally changing the forward computation in the right direction using derivatives.
- Supports interactions otherwise not possible
- Produces smooth animations