Chapter 3: GAs in Scientific Models

• Pro: Computers can simulate evolution
• Con: The simulation is a simplification

Section 3.1: Modelling Interaction Between Learning and Evolution

Lamarckian Hypothesis

Learned knowledge can be passed genetically.

Baldwin Effect

Individuals who can learn quickly are more likely to reproduce.

Hinton, Nowlan (1987)

• Goal: Demonstrate and measure the Baldwin Effect.
• Approach: Use a GA to evolve the learning capacity for a neural network. The NN has 20 possible connections. Each connection can be present, absent or learnable. There is only one correct network.
• Figure 3.1 and Figure 3.2: Fitness Landscapes.
• GA Algorithm:
  1. Generate a random initial population with 25% 0s, 25% 1s, and 50% *s
  2. Evaluate the fitness of each individual using 1000 learning trials
  3. Form the next generation using fitness based selection with single point crossover and no mutation
  4. Go to step 2 (for 50 generations)
• Figure 3.3: Mean Fitness vs. Number of Generations. This study demonstrates the Baldwin Effect - individuals who can learn quickly are selected to reproduce.
• Figure 3.5: Relative Frequencies of Correct, Incorrect and Undecided Alleles.

Ackley, Littman (1992)

• Problem: Agents move around a grid, trying to survive. The agent has an internal store of energy. When the store is high enough, the agent will either clone itself (with mutation) or mate with a nearby agent.
• Figure 3.6: Agent Architecture. The evaluation network is fixed. The action network is learnable using both backpropagation and reinforcement learning. The reinforcement learning works by trying to correlate evaluations at successive time steps. There are a total of 84 weights, each weight is coded by 4 bits, so the size of the search space is 2³³⁶.
• Figure 3.7: Extinction Time. The study investigates ERL (evolution and learning), L (learning only), E (evolution only), F (fixed random weights) and B (brownian random agents). This figure shows that (1) learning is more important than evolution and (2) it is easier to specify an evaluation function than an action function.
• Figure 3.8 and 3.9: Bit Substitutions per Position per Generation. Figure 3.9 purportedly shows that after agents learn to approach food, they can then hard-code this knowledge.