Bridging the Gap Between Theory and Practice: Benchmarking Transfer Evolutionary Optimization

TL;DR

This work addresses the gap between transfer evolutionary optimization theory and practice by challenging the no free lunch theorem with a practical benchmark suite that embodies Big Volume, Big Variety, and Big Velocity. It introduces three real-world-inspired problems—the Knapsack benchmark, Planar Robotic Arm, and Minimalistic Attacks—paired with multi-to-one and many-to-one transfer scenarios to evaluate diverse TrEO algorithms. Through detailed experimental analyses, it demonstrates how model-free and model-based transfers perform under realistic constraints, highlighting the strengths of methods like AMTEA, sTrEO, and MSSTO in different regimes, and it emphasizes the importance of robust benchmarking for advancing practical TrEO solutions. The benchmark suite aims to guide researchers toward more effective, scalable, and agile transfer optimizers for real-world problems.

Abstract

In recent years, the field of Transfer Evolutionary Optimization (TrEO) has witnessed substantial growth, fueled by the realization of its profound impact on solving complex problems. Numerous algorithms have emerged to address the challenges posed by transferring knowledge between tasks. However, the recently highlighted ``no free lunch theorem'' in transfer optimization clarifies that no single algorithm reigns supreme across diverse problem types. This paper addresses this conundrum by adopting a benchmarking approach to evaluate the performance of various TrEO algorithms in realistic scenarios. Despite the growing methodological focus on transfer optimization, existing benchmark problems often fall short due to inadequate design, predominantly featuring synthetic problems that lack real-world relevance. This paper pioneers a practical TrEO benchmark suite, integrating problems from the literature categorized based on the three essential aspects of Big Source Task-Instances: volume, variety, and velocity. Our primary objective is to provide a comprehensive analysis of existing TrEO algorithms and pave the way for the development of new approaches to tackle practical challenges. By introducing realistic benchmarks that embody the three dimensions of volume, variety, and velocity, we aim to foster a deeper understanding of algorithmic performance in the face of diverse and complex transfer scenarios. This benchmark suite is poised to serve as a valuable resource for researchers, facilitating the refinement and advancement of TrEO algorithms in the pursuit of solving real-world problems.

Figures (18)

• Figure 1: In sequential transfer optimization, knowledge from pre-optimized tasks is extracted and stored in a knowledge base. This accumulated knowledge is later utilized to solve the target task.
• Figure 2: Categorization of our benchmarking problem suite, which comprises three representative problems, namely, the Knapsack Problem, Planar Robotic Arm Problem, and Minimalistic Attacks. These problems exemplify discrete, continuous, and mixed optimization domains, respectively. Remarkably, Knapsack Problems have only the volume characteristic. While, Planar Robotic Arm Problems possess volume and variety features, and Minimalistic Attacks exhibit volume and velocity characteristics.
• Figure 3: Heatmaps of source-target relatedness for diverse combinations of $L$ and $\alpha_{max}$, ranging between (0,$\sqrt{2}$) and (0, 1], respectively, for (a) 10 joints and (b) 20 joints. According to the sidebar on the right which marks the fitness range, the hotter (brighter) cells indicate more relatedness to the target (as their fitness values are greater) whereas the colder (darker) ones show less or no relatedness.
• Figure 4: Minimalistic Attacks: In the original BeamRider keyframe (a), an enemy is directly in front of the agent. At this point, the trained agent fires, resulting in an increase in the game's score. However, if the keyframe is disturbed due to an attack (b), causing the agent to mistakenly perceive a bullet, the trained agent becomes more inclined to move left or right to avoid it. Consequently, this alteration in behavior ultimately leads to a successful attack as the agent's original actions are changed.
• Figure 5: Convergence trends for (a) configuration A, (b) configuration B, and (c) configuration C of multi-to-one scenario in the 0/1 Knapsack Problem with 5000 function evaluations.