MS/OS Embedded Optimization with CMA-ES: A One-Page Concept Guide
🧠 Problem Setting: Flexible Job Shop Scheduling (FJSP)
Objective: Minimize Cmax (makespan) of all jobs.
Challenge: Jobs have multiple operations; each operation has multiple machine choices (MS), and scheduling order (OS) matters.
🌟 Embedding MS/OS into Continuous Space
Key Idea: Convert combinatorial MS/OS decisions into a continuous domain [0,1] to enable CMA-ES to operate.
1. Machine Selection (MS):
Each operation chooses a machine index via a normalized real value
x ∈ [0,1]
.Decode:
machine_idx = int(x * num_choices)
.
2. Operation Sequence (OS):
Random keys in [0,1], sorted to determine execution order.
Final Individual Vector:
[MS_1, MS_2, ..., MS_n, OS_1, OS_2, ..., OS_n] ∈ [0,1]^{2n}
⚖️ CMA-ES Algorithm Flow
ask → evaluate → tell loop:
ask(): CMA samples multiple individuals (solutions) from a multivariate Gaussian θ = (m, σ, C).
evaluate(): Decode each individual into MS/OS choices, simulate the schedule, compute
Cmax
.tell(): Update mean
m
, covarianceC
, and step sizeσ
based on fitness ranking.
⚡ Key Enhancements
Plateau Detection: Monitor history of best Cmax; if improvement stagnates, increase
σ
.Guided Local Search (optional):
Apply
VNS
on MS or OS of critical path.Embed only when plateau is detected.
🔄 Nature of the Search Space
[0,1] Continuous Domain:
MS/OS are embedded in a geometric manifold.
CMA-ES adapts the shape of its distribution to fit the underlying optimization landscape.
Problem Density:
Mk01: low degree of freedom → higher σ needed.
Mk10: high flexibility → smaller σ leads to better convergence.
🔍 Objective Function
Let
J = Jobs, P = Processing Times, M = Machines
x ∈ [0,1]^{2n} ⇒ (MS, OS) ⇒ Schedule ⇒ Cmax
Minimize:
f(x) = Cmax(x) = max_{j ∈ J} C[j][-1]
📊 Summary
CMA-ES can effectively solve FJSP by mapping discrete scheduling problems into a continuous optimization space.
Embedding MS/OS decisions into [0,1] enables natural use of evolutionary strategies.
Search guided by adaptive σ, critical path heuristics, and statistical adaptation.
Keywords: CMA-ES, MS/OS embedding, Random Key, Flexible Job Shop Scheduling, Evolutionary Optimization, Cmax, Plateau Escape, Gaussian Search.
沒有留言:
張貼留言