Thomas Weise (汤卫思)

A professor at Hefei University.

2 minute read

Our paper When Frequency Fitness Assignment Fails: Trapped States in Frequency-Guided Local Search has been accepted at the Genetic and Evolutionary Computation Conference (GECCO’2026) taking place from July 13 to 17, 2026 in San José, Costa Rica.

  • Jiazheng ZENG (曾嘉政), Thomas Weise (汤卫思), Zhize WU (吴志泽), and Markus Wagner: When Frequency Fitness Assignment Fails: Trapped States in Frequency-Guided Local Search. Genetic and Evolutionary Computation Conference (GECCO'2026), July 13-17, 2026, San Jose, Costa Rica. New York, NY, USA: ACM.
    , , , , more infos.

Introduction

Frequency Fitness Assignment (, 频率适应度分配) is an “algorithm plugin” that fundamentally changes how the selection step in works. We already discussed the invariance properties of FFA and how FFA can be plugged into in previous posts.

In the paper, we investigate several objective-guided and FFA-guided optimization methods as well as their hybrids to a wide variety of discrete of benchmark functions. Discrete benchmark functions $f:\mathbb{X}\mapsto\mathbb{N}$ are defined over a search space $\mathbb{X}={0,1}^n$ composed of $n$-dimensional bit strings and compute integer results. There exist very simple benchmarks, like OneMax, where the goal is to maximize the number of 1 bits in the strings. There also are hard benchmarks, like the Trap function, which is similar to OneMax but where the “gradient” is flipped so that it points away from the optimum.

Interesting New Findings

In our new paper, we make some interesting findings: First, FFA is designed to be maximally explorative. We expected that it would produce a search that could never prematurely converge, a search that would keep exploring the search space forever. In the paper, we proof the exact conditions under which FFA can get stuck. We show that this might even happen in a very simple scenario: The , i.e., the with a search operator always flipping exactly one bit, can get stuck on OneMax.

Second, we also tackled the low-autocorrelation binary sequence problem (LABS). This is the $\mathcal{NP}$-hard problem of finding a binary sequence of length $n$ whose autocorrelation is as low as possible. We wanted to use it as a benchmark, but discovered several new best-known solutions for some $n\geq600$.

Finally, in the hybrid algorithms combining FFA- and objective-guided search, we introduced a binary operator, i.e., uniform crossover, to transfer information from the FFA-guided to the objective-guided algorithm strand. We find that this works actually quite well.

All of that can be found in our paper, which we will present this July.

Further Reading

  • Thomas Weise (汤卫思), Zhize WU (吴志泽), Xinlu LI (李新路), Yan CHEN (陈岩), and Jörg Lässig: Frequency Fitness Assignment: Optimization without Bias for Good Solutions can be Efficient. IEEE Transactions on Evolutionary Computation (TEVC) 27(4):980-992. August 2023.
    , , , more infos.
  • Thomas Weise (汤卫思), Zhize WU (吴志泽), Xinlu LI (李新路), and Yan CHEN (陈岩): Frequency Fitness Assignment: Making Optimization Algorithms Invariant under Bijective Transformations of the Objective Function Value. IEEE Transactions on Evolutionary Computation (TEVC) 25(2):307-319. April 2021.
    , , , , more infos.
  • Jiazheng ZENG (曾嘉政), Thomas Weise (汤卫思), Zhize WU (吴志泽), and Markus Wagner: When Frequency Fitness Assignment Fails: Trapped States in Frequency-Guided Local Search. Genetic and Evolutionary Computation Conference (GECCO'2026), July 13-17, 2026, San Jose, Costa Rica. New York, NY, USA: ACM.
    , , , , more infos.

You May Also Enjoy

Limiting the Memory and CPU Cores of a Command

3 minute read

Sometimes we want to execute a command or program that may use too many CPU cores or too much memory and thus potentially crash our system. Instead of waiting for this to happen, we can also limit the number of CPU cores as well as the memory that a process can use/allocate. If the program can smoothly run within such limits, it will complete normally. If it tries to allocate more memory than what we permit, it will most likely simply crash. If we have 8 logical CPU cores, but we permit the program to only use 4, then it can also never happen that the computer becomes too slow because of the computational load. Regardless what happens, our computer will remain fully usable.

Finding Non-ASCII Characters in Text Documents

1 minute read

Sometimes we want to find whether a text documents contains only ASCII characters or whether there is some non-ASCII character in it. You see, ASCII is an old 7 bit-based encoding for text. Hence, it supports a very limited range of different characters and certainly not fancy stuff like “ä” let alone “你好”. Most tools today can understand the Unicode character set encoded with UTF-8, which is more or less compatible to that and can store all such characters. Some software, however, still takes offense if characters appear that are outside the ASCII range, e.g., in names of functions in some programming languages. Other software can deal with Unicode well, but dislikes only certain characters in certain places. If we have some strange LaTeX error when compiling a paper draft, sometimes it can help to check whether we accidentally had maybe a “。” instead of a “.” slip in somewhere.

Converting Vector Graphics Formats

3 minute read

Most graphics that we use in publications or reports are vector graphics, such as SVGs, PDFs, sometimes also WMFs, EMFs, EPSs, or PS files. It is often necessary to convert between them, most often between PDF and SVG. This can be done via Inkscape. Here I provide the little script vec2vec.sh, which does this in the terminal. It basically just executes Inkscape with the correct parameters.