Multiobjective Firefly Algorithm (MOFA)

• MOFA is a metaheuristic technique that extends the Firefly Algorithm using Pareto dominance to directly generate diverse Pareto fronts.
• It adapts firefly behavior through weighted aggregation and random walks to efficiently explore continuous optimization problems with nonlinear constraints.
• Benchmark studies indicate MOFA achieves superior convergence metrics compared to NSGA-II, MODE, and other conventional algorithms on standard test and real-world engineering problems.

The Multiobjective Firefly Algorithm (MOFA) is a metaheuristic optimization method for solving continuous multiobjective optimization problems with nonlinear constraints and large parameter spaces. MOFA adapts the principles of the Firefly Algorithm (FA) to directly generate Pareto fronts using Pareto dominance and adaptive randomization, with demonstrated convergence and robustness on benchmark and engineering design problems (Yang, 2013).

1. Formal Problem Definition

MOFA addresses continuous multiobjective problems of the form: $$\begin{aligned} &\min_{x\in\mathbb{R}^d}\; \bigl(f_1(x),\,f_2(x),\,\dots,f_K(x)\bigr),\ &\text{subject to}\quad g_j(x)\le 0,\quad j=1,\dots,J,\ &\qquad\qquad\qquad L_b\le x\le U_b, \end{aligned}$$ where $x\in\mathbb{R}^d$ is the decision vector, $f_k(x)$ denotes the $k$-th objective (to be minimized), $g_j(x)$ are inequality constraints, and $[L_b, U_b]$ are the variable bounds. Each objective may be nonlinear; constraint handling is intrinsic to solution updates.

2. Foundation: The Firefly Algorithm

The original Firefly Algorithm (FA) draws from three principles: unisex attraction among fireflies, attractiveness proportional to perceived brightness which diminishes with distance, and movement of less bright fireflies toward brighter ones. For maximization of a single objective $f(x)$, brightness is identified as $I(x)\propto f(x)$. The FA dynamics are governed by:

• Attractiveness Function:

$\beta(r) = \beta_0\,e^{-\gamma\,r^2}$

where $\beta_0 > 0$ is the base attractiveness, $\gamma \geq 0$ is the light absorption coefficient, and $r = \lVert x_i - x_j \rVert$ is the Euclidean distance.

• Firefly Position Update:

$$x_i^{\,t+1} = x_i^t + \beta_0\,e^{-\gamma\,r_{ij}^2}(x_j^t - x_i^t) + \alpha_t\,\varepsilon_i^t$$

with $\alpha_t$ the randomization parameter (often geometrically decayed), and $\varepsilon_i^t$ a vector of i.i.d. random values (uniform or Gaussian). If $\beta_0=0$, the update degenerates to an unbiased random walk.

3. MOFA Algorithmic Enhancements

MOFA introduces multiobjective extensions centered on Pareto optimality and diversity maintenance:

• Pareto Dominance: For fireflies $i$ and $j$, with objective vectors $\mathbf{f}_i, \mathbf{f}_j$, $j$ dominates $i$ ($\mathbf{f}_j \prec \mathbf{f}_i$) if

$$(\forall k,\; f_k(x_j) \leq f_k(x_i)) \wedge (\exists k,\; f_k(x_j) < f_k(x_i))$$

Under dominance, $i$ is attracted toward $j$ per standard FA movement.

• Weighted Aggregation for Non-Dominated Solutions: If $i$ is non-dominated, a random weight vector $w = (w_1, \ldots, w_K)$, $w_k \geq 0$, $\sum w_k = 1$ is generated, forming a scalarized objective:

$\psi(x) = \sum_{k=1}^K w_k f_k(x)$

The current solution $g^*_t = \arg\min_{x\in\text{pop}} \psi(x)$ is found, and $i$ executes a random walk around $g^*_t$:

$x_i^{t+1} = g^*_t + \alpha_t \varepsilon_i^t$

• Archiving and Diversity: All non-dominated solutions are archived each iteration. The stochasticity in the weight vector encourages coverage of diverse Pareto front regions.

4. MOFA Operational Workflow

The MOFA procedure for $n$ fireflies and $T$ generations, with FA parameters $\beta_0, \gamma, \alpha_0$ is as follows:

1. Initialize $n$ fireflies uniformly within $[L_b, U_b]$.
2. Evaluate all objectives $f_k(x_i)$; identify initial non-dominated set (Pareto front archive, PF).
3. For $t = 1$ to $T$:
   • For each $i=1\ldots n$:
     • For $j \neq i$ in the population:
       • If $x_j$ dominates $x_i$: move $i$ toward $j$ using FA dynamics, enforce bounds/constraints.
     • If $x_i$ non-dominated: generate random weights $w$, compute $g^*_t$ (weighted best), random walk around $g^*_t$, enforce bounds/constraints.
   • Update PF with all non-dominated $x_i$, decay $\alpha_t \leftarrow \alpha_0 0.9^t$.
4. Return final PF as approximate Pareto front.

5. Parameterization and Implementation Guidelines

Extensive parametric analysis yields the following recommendations:

• Population size: $n \approx 50$
• Attractiveness: $\beta_0 \in [0.7, 1.0]$
• Light absorption coefficient: $\gamma \approx 1$, aligning $1/\sqrt{\gamma}$ to problem variable scale
• Randomization: $\alpha_0 \in [0.1, 0.5]$, with $\alpha_t = \alpha_0 \times 0.9^t$

These values ensure sufficient exploration and exploitation for standard engineering test problems.

6. Benchmarks, Performance, and Comparative Evaluation

MOFA has been validated against standard multiobjective benchmark test functions:

Problem	$f_1(x)$, $f_2(x)$ Definition and Domain
SCH	$f_1(x)=x^2$, $f_2(x)=(x-2)^2$, $x \in [-10^3, 10^3]$
ZDT1	$f_1(x)=x_1$, $f_2(x)=g(x)(1-\sqrt{f_1/g(x)})$
ZDT2	$g(x)$ as ZDT1, $f_2(x)=g(x)(1-f_1/g(x))^2$
ZDT3	$$f_2(x)=g(x)[1-\sqrt{f_1/g(x)}-f_1/g(x)\sin(10\pi f_1)]$$
LZ	As defined in the problem statement (uses indexed sums)

Convergence is measured by the generalized distance metric: $$D_g = \frac{1}{N}\sqrt{\sum_{j=1}^N (f_j^e - f_j^t)^2}$$ Lower $D_g$ implies closer approximation to the true Pareto front.

Method	ZDT1	ZDT2	ZDT3	SCH	LZ
VEGA	3.79E–2	2.37E–3	3.29E–1	6.98E–2	1.47E–3
NSGA-II	3.33E–2	7.24E–2	1.14E–1	5.73E–3	2.77E–2
MODE	5.80E–3	5.50E–3	2.15E–2	9.32E–4	3.19E–3
DEMO	1.08E–3	7.55E–4	1.18E–3	1.79E–4	1.40E–3
Bees	2.40E–2	1.69E–2	1.91E–1	1.25E–2	1.88E–2
SPEA	1.78E–3	1.34E–3	4.75E–2	5.17E–3	1.92E–3
MOFA	1.90E–4	1.52E–4	1.97E–4	4.55E–6	8.70E–4

Against NSGA-II, VEGA, MODE, DEMO, Bees, and SPEA, MOFA attained the lowest $D_g$ on all benchmarks after 500 generations ($n=50$). MOFA convergence is exponential (cf. Figure 1 in (Yang, 2013)) and ranked consistently superior.

On real-world design problems (e.g., welded beam and disc-brake), MOFA produced smoother and more complete Pareto fronts in fewer generations than comparators. Computational complexity per iteration is comparable to other population-based approaches, with dominance checks and weight vector randomization adding negligible overhead.

7. Research Directions and Prospective Enhancements

The development and empirical performance of MOFA indicate efficacy on continuous, nonlinear, constrained multiobjective tasks. Targeted research avenues include:

• Theoretical convergence analysis
• Systematic parameter sensitivity investigation
• Hybridization with other metaheuristics
• Adaptation to discrete/combinatorial domains
• Integration of advanced archiving and diversity-maintenance, e.g., hypervolume contribution sorting

These directions address foundational limitations and facilitate broader applicability of MOFA in multiobjective optimization contexts (Yang, 2013).