You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The CRLS and CGLS variants are the ones solving more problems, and even though the difference is rather small the CGLS variant is consistently faster which seems to indicate that it is the most appropriate subsolver for TRUNK.
188
-
The size of the problems were rather small here, so this should be confirmed on larger instance.
204
+
The size of the problems was rather small here, so this should be confirmed on larger instances.
189
205
Moreover, the results may vary depending on the origin of the test problems.
This tutorial is essentially a collection of examples.
15
+
This tutorial demonstrates how to use BenchmarkProfiles.jl to visualize and compare solver performance across multiple test problems.
15
16
16
17
## Performance Profile
17
18
18
-
Performance profiles are straightforward to use. The input is a matrix `T` with entries `T[i,j]` indicating the cost to solve problem `i` using solver `j`. Cost can be, for instance, elapsed time, or number of evaluations. The cost should be positive. If any cost is zero, all measures will be shifted by 1.
19
+
Performance profiles, introduced by Dolan and Moré (2002), provide a graphical way to compare the performance of multiple solvers across a test set. They show the fraction of problems solved by each solver as a function of a performance tolerance.
20
+
21
+
### Understanding Performance Profiles
22
+
23
+
The input is a matrix `T` with entries `T[i,j]` indicating the cost to solve problem `i` using solver `j`. Cost can be, for instance, elapsed time, number of iterations, or function evaluations. The cost should be positive. If any cost is zero, all measures will be shifted by 1.
24
+
25
+
The performance profile plots:
26
+
-**x-axis (τ)**: Performance ratio - how much slower a solver is compared to the best solver for each problem. τ=1 means the solver was fastest, τ=2 means it took twice as long as the fastest solver.
27
+
-**y-axis (ρ(τ))**: Fraction of problems solved within the performance ratio τ. ρ(2) = 0.8 means the solver solved 80% of problems within twice the time of the best solver.
28
+
29
+
**Key interpretations**:
30
+
- The height at τ=1 (left side) shows the fraction of problems where the solver was fastest
31
+
- The right-side height (as τ→∞) shows the fraction of problems successfully solved (robustness)
32
+
- Higher curves are better - the solver solves more problems with smaller performance ratios
`Plots` arguments can be passed to `performance_profile()` or used as they normally would be with `Plots`.
58
74
In the example below, we pass `xlabel` to `performance_profile` and set `ylabel` through `ylabel!`.
59
75
76
+
Common customization options:
77
+
-`lw`: Line width
78
+
-`c` or `color`: Line colors
79
+
-`linestyles`: Line styles (`:solid`, `:dash`, `:dot`, etc.)
80
+
-`xlabel`, `ylabel`: Axis labels
81
+
-`title`: Plot title
82
+
-`legend`: Legend position (e.g., `:bottomright`, `:topleft`)
83
+
60
84
```julia
61
85
using Plots
62
86
@@ -67,3 +91,15 @@ ylabel!("ρ(τ)")
67
91
```
68
92
69
93
94
+
95
+
96
+
97
+
### Additional Parameters
98
+
99
+
The `performance_profile` function accepts several optional keyword arguments:
100
+
101
+
-`logscale::Bool=true`: Use logarithmic scale on the x-axis (default: true). Useful for viewing performance across a wide range of ratios.
102
+
-`sampletol::Number=0`: Tolerance for sampling data points. Can reduce plot complexity for large datasets.
103
+
-`title::String=""`: Title for the plot
104
+
105
+
For more details on performance profiles, see: Dolan, E. D., & Moré, J. J. (2002). Benchmarking optimization software with performance profiles. Mathematical Programming, 91(2), 201-213.
0 commit comments