1+ # ##
2+ # ## Initialization
3+ # ##
4+
5+ # Initialize particle swarm state
6+
7+ function initialize (
8+ r:: Union{ParamRange, Tuple{ParamRange, Any}} ,
9+ tuning:: AbstractParticleSwarm
10+ )
11+ return initialize ([r], tuning)
12+ end
13+
14+ function initialize (rs:: AbstractVector , tuning:: AbstractParticleSwarm )
15+ n = tuning. n_particles
16+ # `length` is 1 for `NumericRange` and the number of categories for `NominalRange`
17+ ranges, parameters, lengths, Xᵢ = zip (_initialize .(tuning. rng, rs, n)... )
18+ indices = _to_indices (lengths)
19+ X = hcat (Xᵢ... )
20+ V = zero (X)
21+ pbest_X = copy (X)
22+ gbest_X = copy (X)
23+ pbest = fill (eltype (X)(Inf ), n)
24+ gbest = similar (pbest)
25+ return ParticleSwarmState (
26+ ranges, parameters, indices, X, V, pbest_X, gbest_X, pbest, gbest
27+ )
28+ end
29+
30+ # Unpack tuple of range and distribution
31+
32+ function _initialize (rng, t:: Tuple{ParamRange, Any} , n)
33+ return _initialize (rng, t[1 ], t[2 ], n)
34+ end
35+
36+ # Initialize parameters with default distributions
37+
38+ function _initialize (rng, r:: NominalRange{T, N} , n) where {T, N}
39+ d = Dirichlet (ones (N))
40+ return _initialize (rng, r, d, n)
41+ end
42+
43+ function _initialize (rng, r:: NumericRange , n)
44+ D = _initializer (MLJTuning. boundedness (r))
45+ return _initialize (rng, r, D, n)
46+ end
47+
48+ _initializer (:: Type{MLJBase.Bounded} ) = Uniform
49+
50+ _initializer (:: Type{MLJTuning.PositiveUnbounded} ) = Gamma
51+
52+ _initializer (:: Type{MLJTuning.Other} ) = Normal
53+
54+ # Fit distributions and initialize parameters
55+
56+ function _initialize (rng, r:: ParamRange , D:: Type{<:Distribution} , n)
57+ throw (ArgumentError (" $D distribution is unsupported for $(typeof (r)) ."
58+ end
59+
60+ function _initialize (rng, r:: NumericRange , D:: Type{<:UnivariateDistribution} , n)
61+ d = Distributions. fit (D, r)
62+ return _initialize (rng, r, d, n)
63+ end
64+
65+ # Initialize parameters with fitted/provided distributions
66+
67+ function _initialize (rng, r:: ParamRange , d:: Distribution , n)
68+ throw (ArgumentError (" $(typeof (d)) distribution is unsupported for $(typeof (r)) ."
69+ end
70+
71+ function _initialize (rng, r:: NominalRange{T, N} , d:: Dirichlet , n) where {T, N}
72+ N != d. alpha0 &&
73+ throw (ArgumentError (" Provided distribution's number of categories don't match $r ."
74+ p = Vector {T} (undef, n)
75+ X = rand (rng, d, n)'
76+ return r, p, N, X
77+ end
78+
79+ function _initialize (rng, r:: NumericRange{T} , d:: UnivariateDistribution , n) where {T}
80+ p = Vector {T} (undef, n)
81+ X = rand (rng, d, n)
82+ return r, p, 1 , X
83+ end
84+
85+ # Helper function to get ranges' corresponding indices
86+ # E.g., `_to_indices((1, 2, 1, 3))` returns `(1, 2:3, 4, 5:7)`
87+
88+ function _to_indices (lengths)
89+ curr = 1
90+ return map (lengths) do length
91+ start = curr
92+ stop = start + length - 1
93+ curr = stop + 1
94+ start == stop ? stop : (start: stop)
95+ end
96+ end
97+
98+ # ##
99+ # ## Retrieval
100+ # ##
101+
102+ # For updating `state.parameters`, the model hyperparameters to be returned, from their
103+ # internal representation `state.X`
104+
105+ function retrieve! (state:: ParticleSwarmState , tuning:: AbstractParticleSwarm )
106+ ranges, params, indices, X = state. ranges, state. parameters, state. indices, state. X
107+ rng = tuning. rng
108+ for (r, p, i) in zip (ranges, params, indices)
109+ _retrieve! (rng, p, r, view (X, :, i))
110+ end
111+ return state
112+ end
113+
114+ function _retrieve! (rng, p, r:: NominalRange , X)
115+ return p .= getindex .(
116+ Ref (r. values),
117+ rand .(rng, Categorical .(X[i,:] for i in axes (X, 1 )))
118+ )
119+ end
120+
121+ function _retrieve! (rng, p, r:: NumericRange{T} , X) where {T<: Integer }
122+ return @. p = round (T, _transform (r. scale, X))
123+ end
124+
125+ function _retrieve! (rng, p, r:: NumericRange{T} , X) where {T<: Real }
126+ return @. p = _transform (r. scale, X)
127+ end
128+
129+ _transform (:: Symbol , X) = X
130+
131+ _transform (scale, X) = scale (X)
0 commit comments