150-word summary of the research
Investigating Adaptive Mutation Rates in Genetic Neural Networks
This experiment investigates the efficiency of evolutionary algorithms in finding Nash Equilibrium in a competitive multi-agent environment (Tag). We compare a standard static mutation rate against a novel adaptive mutation strategy where the mutation rate scales inversely with an agent's fitness score. The hypothesis is that adaptive mutation—mimicking biological 'stress-induced mutagenesis'—will allow low-fitness populations to explore the solution space aggressively while high-fitness populations exploit their successful strategies, resulting in significantly faster convergence to a stable strategy (Nash Equilibrium).
Introduction and background research
Deep Reinforcement Learning (DRL) is computationally expensive and hyperparameter-sensitive. Simple evolutionary algorithms are robust but often slow to converge because a fixed mutation rate is inefficient: too high disrupts good policies, too low causes stagnation. finding the optimal balance is difficult.
A stable state in a game where no player can improve their outcome by unilaterally changing their strategy.
Dynamically adjusting the rate of genetic change based on performance (fitness). High stress (low fitness) = High mutation.
A measure of the randomness of an agent's actions. High entropy = exploration; Low entropy = exploitation/convergence.
The testable prediction guiding this research
the mutation rate of a neural network is dynamically scaled inversely to its fitness score,
the population will reach a state of Policy Entropy stability (Nash Equilibrium) in fewer generations than a control group with a static mutation rate,
this mechanism mimics biological 'stress-induced mutagenesis,' allowing poor-performing agents to explore the solution space rapidly while high-performing agents preserve their successful traits, balancing exploration and exploitation more efficiently.
The experimental design controls all factors except the independent variable
The factor that is intentionally changed
The outcomes being measured
Factors kept constant to ensure validity