Add JSO-compliant solver documentation [draft]

docs/make.jl

@@ -9,7 +9,11 @@ makedocs(
     prettyurls = get(ENV, "CI", nothing) == "true",
   ),
   sitename = "SolverCore.jl",
-  pages = ["Home" => "index.md", "Reference" => "reference.md"],
+  pages = [
+    "Home" => "index.md",
+    "JSO-compliant solvers" => "jso-compliant.md",
+    "Reference" => "reference.md",
+  ],
 )
 
 deploydocs(

docs/src/jso-compliant.md

@@ -0,0 +1,32 @@
+# JSO-compliant solvers
+
+The following are JSO-compliance expectations and recommendations that are implemented in JuliaSmoothOptimizers.
+
+## Mandatory
+
+- Create a function `solver_name(nlp::AbstractNLPModel; kwargs...)` that returns a `GenericExecutionStats`.
+
+## Recommended
+
+- Create a solver object `SolverName <: SolverCore.AbstractOptimizationSolver`.
+- Store all memory-allocating things inside this object. This includes vectors, matrices, and possibly other objects.
+- One of these objects should be `solver.x`, that stores the current iterate.
+- Implement a constructor `SolverName(nlp; kwargs...)`.
+- Implement `SolverCore.solve!(solver, nlp::AbstractNLPModel, stats::GenericExecutionStats)` and change `solver_name` to create a `SolverName` object and call `solve!`.
+- Make sure that `solve!` is not allocating.
+- Accept the following keyword arguments (`T` is float type and `V` is the container type):
+  - `x::V = nlp.meta.x0`: The starting point.
+  - `atol::T = sqrt(eps(T))`: Absolute tolerance for the gradient. Use in conjunction with the relative tolerance below to check $\Vert \nabla f(x_k)\Vert \leq \epsilon_a + \epsilon_r\Vert \nabla f(x_0)\Vert$.
+  - `rtol::T = sqrt(eps(T))`: Relative tolerance for the gradient. See `atol` above.
+  - `max_eval::Int = -1`: Maximum number of objective function evaluation plus constraint function evaluations. Negative number means unlimited.
+  - `max_iter::Int = typemax(Int)`: Maximum number of iterations.
+  - `max_time::Float64 = 30.0`: Maximum elapsed time.
+  - `verbose::Int = 0`: Verbosity level. `0` means nothing and `1` means something. There are no rules on the level of verbosity yet.
+  - `callback = (nlp, solver, stats) -> nothing`: A callback function to be called at the end of an iteration, before exit status are defined.
+- Use `set_status!` and `get_status` to update `stats` before starting the method loop, and at the end of every iteration.
+- Call the `callback` after running `set_status!` in both places.
+- Define `done = stats.status != :unknown` and loop with `while !done`.
+- To check for logic errors and stop the method use `set_status!(stats, ...)`, `done = true`, and `continue`, where the second argument of `set_status!` is one of the statuses available in SolverCore.STATUSES. You can call `SolverCore.show_statuses()` to see them. If you need more specific statuses, create an issue.
+- Use the `set_...!(stats, ...)` functions from `SolverCore` to update the `stats`. For instance, `set_objective!(stats, f)`, `set_time!(stats, time() - start_time)`, and `set_dual_residual!(stats, gnorm)`.
+- Don't log when `verbose == 0`. When logging, use `@info log_header(...)` and `@info log_row(...)`.
+- Write docstrings for `SolverName`. The format is still a bit loose.