Real-valued Coded Evolutionary Algorithms

1 Binary Coded Genetic Algorithms

2 Key Concepts in Binary Coding

• schema: a template that identifies a subset of streams with similarities at certain string positions.
  一个长度为L的chromosome，包含了3^L个schemata (0, 1 or * at each position)。
  而一个大小为M的population，我们需要评估至多M*3^L个schemata
• implicit parallelism: to find fitter schemata, and this manipulation is based on M·3^L schemata. (broaden search space)

3 Drawbacks of Binary Coding

• 【1】Hamming cliff problem
  • 汉明距离：表示两个（相同长度）字符串对应位置的不同字符的数量
  • 汉明悬崖问题：二进制和十进制，000跟100只翻转了一个(one-bit mutation)但相差4 (large jump)，011和100翻转了3个(multi-bit mutation)但只相差1 (small jump)。
  • 换言之，the change in genotypic space does not reflec the change in phenotypic space.
• Solution to Hamming cliff problem: Gray encoding
  • 最左边是standard binary code，往右一直异或，取值0或1
  • equation as follow
    
• 【2】problem in discrete search spaces: redundancy problem
  • redundancy problem: when the variables belongs to a finite discrete set with a cardinal different from a power of two, some binary strings are redundant, which correspond infeasible solutions.
  • 举个例子：我有一个集合A = {0, 2, 3}，the cardinal of this set is 3。我们需要一个长度为2 bits的binary string去表示它 (encode 0 1 2 3)。其中，1是infeasible的，这就是redundency。
• 【3】problem in continuous search spaces: precision problem
  • 从01字符串，decode到real value，想要约精确就需要更长的字符串
  • 换言之，the precision depends on length L
• Solution to continuous optimisation problems: Real-valued vector representation
  • no differeneccs between genotypes and phenotypes
  • each element (also called gene) in this vector is a real value, and that represents a real value variable of this problem.
  • 不会受到encode和decode的字符长度限制
  • 一些based on real-valued vector representation的evolutioanry algorithms：Evolution Strategies, Evolutionary Programming and Differential Evolution。
  • 总结advantages: (1) natural, simple and faster (no need to encode and decode); (2) better preision and easy to handle large dimensional problems

4 Real-valued Vector Operators

• 回顾real-valued vector representation

  • it’s a solution to continuous optimisation problems
  • it’s simple, natural and faster (no need to encode and decode)
  • and it has better preision and is easy to handle large dimensional problems

• Real value mutation

  • 根据概率p_m去randomly select a parent for mutation, then randomly select a gene c_i from the parent
  • different forms of mutation operators
    • Uniform mutation: 在variable x_i的intervel bound里，uniformly generate一个数，并用它去替换gene c_i。
    • Gaussian mutation: 在以gene c_i为mean的高斯分布中（标准差根据interval bound决定）计算一个数，去替换gene c_i。里面的min, max function是为了确保计算出来的数字不要超出interva bound。
      
    • Non-uniform mutation:
      • v_i是upper bound，u_i是lower bound
      • ❓ 问题：为什么这个delta的公式是这样的？What’s ths intuition?
        ➡ 如果t/g_m小，则Δ大。that means mutation at the beginning of generation would give us large variation (large standard diviation), and that helps the exploration at the beginning. After a few generations, mutation would be smaller, and that is exploitation.

• Real-valued crossover

  • 选择two parents, then apply a corssover operator
  • 一些crossover operators

    • Flat crossover: 两个父母中，同一个位置的gene随机在interval中选一个value
      

    • Simple crossover: 随机一个crossover point，在这个点以后的所有gene全部swap，这样的到了两个new offsprings
      

    • Whole arithemetical crossover: 根据公式计算新的两个offsprings
      

    • Local arithmetical crossover: 同上，但此处的α不是一个数，是一个Vector
      

    • Single arithmetical crossover: 只改变一个gene，其他直接复制parents
      

    • BLX-α crossover: 计算得到一个new offspring
      

⚠ 不论是mutation operator还是crossover operator，都可以从最简单开始尝试。