Add benchmark for observable functions

benchmarking/ObservableFunctions/benchmark_script.jl

@@ -0,0 +1,232 @@
+using AbstractAlgebra
+using DataStructures
+using StructuralIdentifiability
+using StructuralIdentifiability: lie_derivatives_up_to
+using RationalFunctionFields
+using PrettyTables
+using BenchmarkTools
+
+# -----------------------------------------------------------------------------
+# This script benchmarks StructuralIdentifiability.jl's algorithm for computing
+# identifiable functions against a "naive" baseline that expands Lie derivatives
+# up to a brute-force-determined order and then simplifies the resulting
+# rational function field directly. It reports, per model, the prolongation
+# orders/degree/size needed by each approach and the resulting runtimes.
+# -----------------------------------------------------------------------------
+
+"""
+    prolongation_dict(ode, ords)
+
+Build the `Dict` mapping each output equation `ode.y_equations[y]` to its
+prolongation order `ords[i]`, as expected by `lie_derivatives_up_to`.
+"""
+function prolongation_dict(ode, ords)
+    return Dict(ode.y_equations[y] => ords[i] for (i, y) in enumerate(ode.y_vars))
+end
+
+"""
+    get_prolongation_orders(ode, with_identifiable_functions) -> Vector{Int}
+
+For each output of `ode`, find the smallest prolongation order (number of Lie derivatives) needed to
+get the full field of observable functions.
+
+If `with_identifiable_functions`, the computations strats with the field of 
+identifiable functions in parameter.
+
+Returns a vector of orders, one per entry of `ode.y_vars`.
+"""
+function get_prolongation_orders(ode, with_identifiable_functions)
+    orders = [0 for _ in ode.y_vars]
+    original_gens = [values(ode.y_equations)...]
+    if with_identifiable_functions
+        original_gens = vcat(original_gens, find_identifiable_functions(ode))
+    end
+    rff = RationalFunctionField(original_gens)
+    curr_trdeg = trdeg(rff)
+
+    # Independently grow the order for each output until the transcendence
+    # degree of the accumulated field saturates.
+    for i in 1:length(ode.y_vars)
+        while true
+            orders[i] += 1
+            original_gens = vcat(original_gens, lie_derivatives_up_to(ode, prolongation_dict(ode, orders)))
+            new_trdeg = trdeg(RationalFunctionField(original_gens))
+            if new_trdeg == curr_trdeg
+                break
+            else
+                curr_trdeg = new_trdeg
+            end
+        end
+    end
+    return orders
+end
+
+"""
+    trdeg(rff) -> Int
+
+Compute the transcendence degree of a `RationalFunctionField`
+"""
+function trdeg(rff)
+    RationalFunctionFields.update_trbasis_info!(rff, 0.99)
+    return length(rff.trbasis)
+end
+
+function max_over_list(funcs, property)
+    return maximum(f -> property(denominator(f)) + property(numerator(f)), funcs)
+end
+
+"""
+    get_deg_size(ode, ords) -> (deg, size)
+
+Compute the Lie-derivative prolongations of `ode` up to orders `ords`
+(one per output, see `prolongation_dict`) and return the worst-case total
+degree and worst-case size (`deg`, `size`) among them.
+"""
+function get_deg_size(ode, ords)
+    prolongations = lie_derivatives_up_to(ode, prolongation_dict(ode, ords))
+    deg = max_over_list(prolongations, total_degree)
+    size = max_over_list(prolongations, length)
+    return (deg, size)
+end
+
+"""
+    get_orders_and_degrees(ode) -> Vector
+
+Summarize the prolongation cost of `ode` for the benchmarking table. Computes:
+- `ord_low`/`deg_low`/`size_low`: prolongation orders (without identifiable
+  functions as extra generators) and the resulting degree/size;
+- `ord_high`/`deg_high`/`size_high`: the same but with identifiable functions
+  included as extra generators;
+- `deg_final`/`size_final`: degree/size of the model's identifiable functions
+  themselves (including states), as returned by the actual algorithm.
+
+Returns a flat vector in that order, ready to be stacked into a table row.
+"""
+function get_orders_and_degrees(ode)
+    ord_low = get_prolongation_orders(ode, false)
+    ord_high = get_prolongation_orders(ode, true)
+
+    id_funcs = find_identifiable_functions(ode, with_states = true)
+
+    return [
+        ord_low, get_deg_size(ode, ord_low)...,
+        ord_high, get_deg_size(ode, ord_high)...,
+        max_over_list(id_funcs, total_degree),
+        max_over_list(id_funcs, length),
+    ]
+end
+
+"""
+    compute_direct(ode)
+
+Direct approach for computing observable functions: expand Lie derivatives
+up to the orders found by `get_prolongation_orders(ode, false)`, then
+directly simplify the generating set of the resulting rational function field.
+"""
+function compute_direct(ode)
+    ords = get_prolongation_orders(ode, false)
+    prolongations = lie_derivatives_up_to(ode, prolongation_dict(ode, ords))
+    return simplified_generating_set(RationalFunctionField(prolongations))
+end
+
+"""
+    get_runtimes(models, to_skip) -> Vector
+
+Benchmark, for each `(name, ode)` pair in `models`, the runtime of
+`find_identifiable_functions` (the algorithm) versus `compute_direct`. 
+Models listed in `to_skip` (by name) skip the naive benchmark
+(e.g. because it is prohibitively slow) and are recorded with `Inf` time and
+`Inf` speedup instead.
+
+Returns a vector of rows `[name, time_algo, time_naive, speedup]`, where
+`speedup = time_naive / time_algo`.
+"""
+function get_runtimes(models, to_skip)
+    result = []
+    for (name, ode) in models
+        time_algo = @belapsed find_identifiable_functions($ode, with_states = true)
+        time_naive = Inf
+        if !(name in to_skip)
+            time_naive = @belapsed compute_direct($ode)
+        end
+        push!(result, [name, time_algo, time_naive, time_naive / time_algo])
+    end
+    return result
+end
+
+###########################
+# Example models
+###########################
+
+siwr = @ODEmodel(
+    S'(t) = -beta_W * S(t) * W(t) - beta_I * S(t) * I(t),
+    I'(t) = beta_W * S(t) * W(t) + beta_I * S(t) * I(t) - gamma * I(t),
+    W'(t) = alpha * I(t) - zeta * W(t),
+    y(t) = W(t)
+)
+
+MM = @ODEmodel(
+    x1'(t) = V1 * x1(t) / (1 + K1 * x1(t) * (1 + I(t) / L1) + K2 * x2(t) * (1 + I(t) / L2)),
+    x2'(t) = V2 * x2(t) / (1 + K1 * x1(t) * (1 + I(t) / L1) + K2 * x2(t) * (1 + I(t) / L2)),
+    y(t) = x1(t) + x2(t)
+)
+
+sliqr = @ODEmodel(
+    S'(t) = -b * In(t) * S(t) / N - u(t) * S(t) / N,
+    L'(t) = b * In(t) * S(t) / N - a * L(t),
+    In'(t) = a * L(t) - g * In(t) + s * Q(t),
+    Q'(t) = (1 - e) * g * In(t) - s * Q(t),
+    y(t) = In(t) / N
+)
+
+cancer1 = @ODEmodel(
+    x'(t) = mu_m * (1 - q / Q(t)) * x(t) - (v(t) * R / (R + Q(t)) + delta * x(t)) * x(t),
+    v'(t) = -d * v(t),
+    Q'(t) = (g1 * u(t) + g2) * (Q_m - Q(t)) - mu_m * (Q(t) - q),
+    P'(t) = b * Q(t) + s * x(t) * Q(t) - e * P(t),
+    y(t) = P(t), y2(t) = Q(t)
+)
+
+cancer2 = @ODEmodel(
+    x'(t) = mu_m * (1 - q / Q(t)) * x(t) - (v(t) * R / (R + Q(t)) + delta * x(t)) * x(t),
+    v'(t) = -d * v(t),
+    Q'(t) = (g1 * u(t) + g2) * (Q_m - Q(t)) - mu_m * (Q(t) - q),
+    P'(t) = b * Q(t) + s * x(t) * Q(t) - e * P(t),
+    y(t) = P(t), y2(t) = v(t)
+)
+
+eaihrd = @ODEmodel(
+    A'(t) = a * E(t) - r1 * A(t),
+    I'(t) = s * E(t) - (h + r2) * I(t),
+    H'(t) = h * I(t) - (r3 + d) * H(t),
+    R'(t) = r1 * A(t) + r2 * I(t) + r3 * H(t),
+    D'(t) = d * H(t),
+    E'(t) =
+        (N - A(t) - I(t) - H(t) - R(t) - D(t) - E(t)) * (c1 * A(t) + c2 * I(t)) - (a + s) * E(t),
+    y(t) = D(t)
+)
+
+models = OrderedDict(
+    :SIWR => siwr,
+    :MM => MM,
+    :SLIQR => sliqr,
+    :cancer1 => cancer1,
+    :cancer2 => cancer2,
+    :EAIHRD => eaihrd
+)
+
+###########################
+# Run benchmarks and report
+###########################
+
+# Table 1: prolongation orders and degree/size, with and without identifiable
+# functions as extra generators, plus the degree/size of the final identifiable
+# functions themselves.
+degsize_data = stack([[name, get_orders_and_degrees(ode)...] for (name, ode) in models]; dims = 1)
+
+pretty_table(degsize_data; column_labels = ["Name", "OrdLow", "DegLow", "SizeLow", "OrdHigh", "DegHigh", "SizeHigh", "DegFinal", "SizeFinal"])
+
+# Table 2: runtime of the algorithm vs. the direct baseline (EAIHRD is skipped
+# for the naive computation as it is too slow).
+runtimes = get_runtimes(models, [:EAIHRD])
+pretty_table(stack(runtimes; dims = 1); column_labels = ["Name", "Time algo", "Time naive", "Speedup"])