1+ """
2+ AdaptiveParticleSwarm(n_particles = 3,
3+ c1 = 2.0,
4+ c2 = 2.0,
5+ prob_shift = 0.25,
6+ rng = Random.GLOBAL_RNG)
7+
8+ Instantiate an adaptive particle swarm optimization tuning strategy. A swarm is
9+ initiated by sampling hyperparameters with their customizable priors, and new
10+ models are generated by referencing each member's and the swarm's best models so
11+ far.
12+
13+ ### Supported Ranges and Discrete Hyperparameter Handling
14+
15+ See [`ParticleSwarm`](@ref) for more information about supported ranges and how
16+ discrete hyperparameters are handled.
17+
18+ ### Algorithm
19+
20+ Hyperparameter ranges are sampled and concatenated into position vectors for
21+ each swarm particle. Velocity is initiated to be zeros, and in each iteration,
22+ every particle's position is updated to approach its personal best and the
23+ swarm's best models so far with the equations:
24+
25+ \$ vₖ₊₁ = w⋅vₖ + c₁⋅rand()⋅(pbest - xₖ) + c₂⋅rand()⋅(gbest - xₖ)\$
26+
27+ \$ xₖ₊₁ = xₖ + vₖ₊₁\$
28+
29+ Coefficients `w`, `c1`, `c2` are adaptively adjusted at each iteration by
30+ determining the evolutionary phase of the swarm. We calculate the evolutionary
31+ factor by comparing the mean distance from each particle to other members of the
32+ swarm. This factor is then used to classify whether the swarm is in exploration,
33+ exploitation, convergence, or jumping out phase and calibrate the tuning
34+ hyperparameters accordingly. For more information, refer to "Adaptive Particle
35+ Swarm Optimiztion" by Zhan, Zhang, Li, and Chung. Note that we omit the elitist
36+ learning strategy in the paper.
37+
38+ New models are then generated for evaluation by mutating the fields of a deep
39+ copy of `model`. If the corresponding range has a specified `scale` function,
40+ then the transformation is applied before the hyperparameter is returned. If
41+ `scale` is a symbol (eg, `:log`), it is ignored.
42+ """
43+ mutable struct AdaptiveParticleSwarm{R<: AbstractRNG } <: AbstractParticleSwarm
44+ n_particles:: Int
45+ c1:: Float64
46+ c2:: Float64
47+ prob_shift:: Float64
48+ rng:: R
49+ end
50+
51+ # Constructor
52+
53+ function AdaptiveParticleSwarm (;
54+ n_particles= 3 ,
55+ c1= 2.0 ,
56+ c2= 2.0 ,
57+ prob_shift= 0.25 ,
58+ rng:: R = Random. GLOBAL_RNG
59+ ) where {R}
60+ swarm = AdaptiveParticleSwarm {R} (n_particles, c1, c2, prob_shift, rng)
61+ message = MLJTuning. clean! (swarm)
62+ isempty (message) || @warn message
63+ return swarm
64+ end
65+
66+ # Validate tuning hyperparameters
67+
68+ function MLJTuning. clean! (tuning:: AdaptiveParticleSwarm )
69+ warning = " "
70+ if tuning. n_particles < 3
71+ warning *= " AdaptiveParticleSwarm requires at least 3 particles. " *
72+ " Resetting n_particles=3. "
73+ tuning. n_particles = 3
74+ end
75+ c1, c2 = tuning. c1, tuning. c2
76+ if ! (1.5 ≤ c1 ≤ 2.5 ) || ! (1.5 ≤ c2 ≤ 2.5 ) || (c1 + c2 > 4 )
77+ c1, c2 = _clamp_coefficients (c1, c2)
78+ warning *= " AdaptiveParticleSwarm requires 1.5 ≤ c1 ≤ 2.5, 1.5 ≤ c2 ≤ 2.5, and " *
79+ " c1 + c2 ≤ 4. Resetting coefficients c1=$(c1) , c2=$(c2) . "
80+ tuning. c1 = c1
81+ tuning. c2 = c2
82+ end
83+ if ! (0 ≤ tuning. prob_shift < 1 )
84+ warning *= " AdaptiveParticleSwarm requires 0 ≤ prob_shift < 1. " *
85+ " Resetting prob_shift=0.25. "
86+ tuning. prob_shift = 0.25
87+ end
88+ return warning
89+ end
90+
91+ # Helper function to clamp swarm coefficients in the interval [1.5, 2.5] with a sum of less
92+ # than or equal to 4
93+
94+ function _clamp_coefficients (c1, c2)
95+ c1 = min (max (c1, 1.5 ), 2.5 )
96+ c2 = min (max (c2, 1.5 ), 2.5 )
97+ scale = 4. / (c1 + c2)
98+ if scale < 1
99+ c1 *= scale
100+ c2 *= scale
101+ end
102+ return c1, c2
103+ end
104+
105+ # Initial state
106+
107+ function MLJTuning. setup (tuning:: AdaptiveParticleSwarm , model, ranges, n, verbosity)
108+ # state, evolutionary phase, swarm coefficients
109+ return (initialize (ranges, tuning), nothing , tuning. c1, tuning. c2)
110+ end
111+
112+ # New models
113+
114+ function MLJTuning. models (
115+ tuning:: AdaptiveParticleSwarm ,
116+ model,
117+ history,
118+ (state, phase, c1, c2),
119+ n_remaining,
120+ verbosity
121+ )
122+ n_particles = tuning. n_particles
123+ if ! isnothing (history)
124+ sig = MLJTuning. signature (history[1 ]. measure[1 ])
125+ measurements = similar (state. pbest)
126+ map (history[end - n_particles+ 1 : end ]) do h
127+ measurements[h. metadata] = sig * h. measurement[1 ]
128+ end
129+ pbest! (state, measurements, tuning)
130+ gbest! (state)
131+ f = _evolutionary_factor (state. X, argmin (state. pbest))
132+ phase = _evolutionary_phase (f, phase)
133+ w, c1, c2 = _adapt_parameters (tuning. rng, c1, c2, f, phase)
134+ move! (tuning. rng, state, w, c1, c2)
135+ end
136+ retrieve! (state, tuning)
137+ fields = getproperty .(state. ranges, :field )
138+ new_models = map (1 : n_particles) do i
139+ clone = deepcopy (model)
140+ for (field, param) in zip (fields, getindex .(state. parameters, i))
141+ recursive_setproperty! (clone, field, param)
142+ end
143+ (clone, i)
144+ end
145+ return new_models, (state, phase, c1, c2)
146+ end
147+
148+ # Helper function to calculate the evolutionary factor and phase
149+
150+ function _evolutionary_factor (X, gbest_i)
151+ n_particles = size (X, 1 )
152+ dists = zeros (n_particles, n_particles)
153+ for i in 1 : n_particles
154+ for j in i+ 1 : n_particles
155+ dists[j, i] = dists[i, j] = norm (X[i, :] - X[j, :])
156+ end
157+ end
158+ mean_dists = sum (dists, dims= 2 ) / (n_particles - 1 )
159+ min_dist, max_dist = extrema (mean_dists)
160+ gbest_dist = mean_dists[gbest_i]
161+ f = (gbest_dist - min_dist) / max (max_dist - min_dist, sqrt (eps ()))
162+ return f
163+ end
164+
165+ function _evolutionary_phase (f, phase)
166+ # Classify evolutionary phase
167+ μs = [μ₁ (f), μ₂ (f), μ₃ (f), μ₄ (f)]
168+ if phase === nothing # first iteration
169+ phase = argmax (μs)
170+ else
171+ next_phase = mod1 (phase + 1 , 4 )
172+ # switch to next phase if possible
173+ if μs[next_phase] > 0
174+ phase = next_phase
175+ # stay in current phase is possible, else pick the most likely phase
176+ elseif μs[phase] == 0
177+ phase = argmax (μs)
178+ end
179+ end
180+ return phase
181+ end
182+
183+ # Helper functions to calculate probabilities of the four evolutionary states
184+
185+ μ₁ (f) = f ≤ 0.4 ? 0.0 :
186+ f ≤ 0.6 ? 5 * f - 2 :
187+ f ≤ 0.7 ? 1 :
188+ f ≤ 0.8 ? - 10 * f + 8 :
189+ 0.0
190+
191+ μ₂ (f) = f ≤ 0.2 ? 0.0 :
192+ f ≤ 0.3 ? 10 * f - 2 :
193+ f ≤ 0.4 ? 1.0 :
194+ f ≤ 0.6 ? - 5 * f + 3 :
195+ 0.0
196+
197+ μ₃ (f) = f ≤ 0.1 ? 1.0 :
198+ f ≤ 0.3 ? - 5 * f + 1.5 :
199+ 0.0
200+
201+ μ₄ (f) = f ≤ 0.7 ? 0.0 :
202+ f ≤ 0.9 ? 5 * f - 3.5 :
203+ 1.0
204+
205+ # Adaptive control of swarm's parameters
206+
207+ function _adapt_parameters (rng, c1, c2, f, phase)
208+ w = 1.0 / (1.0 + 1.5 * exp (- 2.6 * f)) # update inertia
209+ δ = rand (rng) * 0.05 + 0.05 # coefficient acceleration
210+ if phase === 1 # exploration
211+ c1 += δ
212+ c2 -= δ
213+ elseif phase === 2 # exploitation
214+ δ *= 0.5
215+ c1 += δ
216+ c2 -= δ
217+ elseif phase === 3 # convergence
218+ δ *= 0.5
219+ c1 += δ
220+ c2 += δ
221+ else # jumping out
222+ c1 -= δ
223+ c2 += δ
224+ end
225+ c1, c2 = _clamp_coefficients (c1, c2)
226+ return w, c1, c2
227+ end
0 commit comments