Abstract
This paper introduces SUNA (Spectrum-diverse Unified Neuroevolution Architecture), a framework designed to evolve neural networks by integrating diverse biological features such as varying time scales and neuromodulation. By employing a novel "Spectrum Diversity" metric that evaluates structural characteristics rather than individual genes, SUNA addresses the "curse of dimensionality" in neuroevolution. Experimental results demonstrate that SUNA significantly outperforms the NEAT algorithm on complex, memory-dependent, and logic-based tasks, while requiring fewer hyperparameters and showing greater robustness to unnormalized data.
Introduction
Neuroevolution aims to automate the design of neural networks through evolutionary algorithms. While successful, standard methods like NEAT often rely on simplified neuron models. Biological brains utilize complex mechanisms—including heterogeneous reaction speeds, neuromodulation, and diverse activation functions—to solve real-world problems. Integrating these features into artificial systems typically expands the search space exponentially, leading to the "curse of dimensionality." Additionally, many existing systems require manual data normalization, limiting their autonomy.
Methods
SUNA introduces two primary innovations: a Unified Neuron Model and Spectrum Diversity.
Unified Neuron Model
The Unified Neuron Model extends standard neurons with evolvable properties:
- Adaptation Speed: A parameter allowing neurons to react at different rates, providing built-in memory for temporal data.
- Control Neurons: Special neurons that modulate the activity of others, mimicking biological inhibition and excitation.
Structural evolution in SUNA involves mutations such as adding nodes and connections to the network topology, allowing the system to grow in complexity as needed.
Spectrum Diversity
Instead of tracking individual genes via innovation numbers, SUNA utilizes Spectrum Diversity. This approach characterizes networks by a "spectrum"—a histogram of structural features like neuron types, activation functions, and speeds. A Novelty Map is used to maintain population diversity based on these structural profiles, preventing premature convergence to local optima and protecting innovative structural "flavors."
Recurrent Connections and Memory
For tasks requiring memory (Non-Markovian problems), the evolution of feedback connections is crucial. These recurrent loops, combined with evolvable adaptation speeds, allow SUNA to process time-series data effectively without manual feature engineering.
Evaluation and Results
SUNA was evaluated against NEAT across five classes of problems, including Mountain Car, Double Pole Balancing (Markov and Non-Markov), Multiplexer logic, and Function Approximation.
Key findings include:
- Superior Performance: In the Non-Markov Double Pole Balancing task, SUNA achieved 70% of the maximum reward, nearly doubling NEAT's 37%.
- Robustness: SUNA performs equally well on "raw" unnormalized data, whereas NEAT's performance degrades significantly without precise input scaling.
- Efficiency: Despite its complexity, SUNA requires only 8 hyperparameters, compared to 33 for NEAT, making it significantly easier to tune for new applications.
Discussion and Limitations
Ablation studies revealed that different biological features are specialized for different tasks: neuromodulation was critical for logic (Multiplexer), while slow neurons were essential for memory-dependent tasks.
However, SUNA can be slower to converge on simple Markovian problems where its expanded search space is less advantageous. Future work will explore "Indirect Encoding" (e.g., HyperNEAT) to scale these unified models to larger architectures, potentially enabling the evolution of more complex robotic or visual processing systems.
Summary updated