Lecture 8: Evolutionary Robotics

Teacher: Stéphane Doncieux, Nicolas Bredèche

Random generation ⟶ Evaluation ⟶ Selection ⟶ Variation (Until termination)

⇒ Evolutionary Computation

EvolutionaryAlgorithm():
    t = 0

    RandomInitialization(P_t):
        for i in range(N+1):
            Evaluate(P_i(t))

            while condition_not_met:
                Q = 0

                for i in range(N+1):
                    x = Recombine_select_repro(P_t)
                    x = Mutate(x)
                    Evaluate(x)
                    Q = x ∪ Q
                P(t+1) = Select_replace(P_t, Q)
                t += 1

Algorithms

Microbial GA

Very simple algorithm, very few prerequisites

1. Start from a genotype population
2. Random selection
3. Ranking

4. Modify the losing genotypes:

   • Random recombination: inject a part of the winner in the losers
   • Mutation in the losers
5. Repeat

Efficient parameter optimization: CMA-ES

State-of-the art parameter optimization algorithm

1. Initial sampling ⟶ selection according to fitness
2. Parametric models on this distribution (mean, covariance matrix)
3. Make mutations in the gradient (of the fitness function) direction

Multi-objective optimization: NSGA-II

Multi-objective optimization: no total order

⟹ we focus on the Pareto boundary: solutions that don’t make any compromise (NSGA-II: state-of-the-art)