Pranav Mehrotra

At the algorithmic core, the GA evolves solutions using crossover and mutation, while local refinements exploit Euclidean structure to improve tour quality. To ensure scalability, I implemented a dual-level strategy. Horizontally, an island model runs in Docker containers: each evolves subpopulations independently and periodically exchanges solutions through Redis, where a Python coordinator applies stochastic best-of-K migration before redistributing updated populations. Vertically, within each container, the C++ engine leverages OpenMP parallelism to accelerate crossover, mutation, and refinement across CPU cores.

The systems design ensures modularity and efficiency: C++ delivers raw computational performance, Python orchestrates container lifecycles and migration, Redis serves as a lightweight message broker, and Docker guarantees reproducibility and portability.

This integration allowed the system to achieve near-optimal tours (within ~0.5% error) on benchmarks with up to 1000 cities—a scale at which brute-force search (10^2564 possible tours) is physically impossible. By blending evolutionary computation, heuristic search, parallel computing, and distributed systems, this project demonstrates how hybrid AI + systems engineering can address massive combinatorial problems.

Rather than using backpropagation, I treat the policy network as a black box and optimize directly in weight space. The network parameters are modeled by a block-diagonal CMA-ES: each small group of neurons has its own Gaussian (mean + covariance) over weights and biases. For each episode I sample one full network, run a full episode, and use the total return as the fitness.

I keep a rolling buffer of the last 20 policies, select the top 5 "elite" policies, and update the block-wise Gaussians using CMA-ES style mean and covariance adaptation. This gives a structured, second-order evolutionary update rule over the policy, with no gradients, no TD-learning, and no backprop.

On the systems side, I built a Redis-backed Gymnasium server in Python, a C++ CMA-ES agent implementing this block-wise + FIFO + top-k update, and a set of baseline agents (PPO, A2C, SAC, DQN via Stable-Baselines3) that share the same Redis interface. This lets me benchmark SAGEnet directly against modern RL algorithms and study when evolutionary policy search can realistically act as an alternative to gradient-based policy optimization.

Optimized reductions using threadgroups, grouped work sizes, and persistent GPU buffers to minimize allocations and kernel launches, beating NumPy on large (> 5 × 10^6 element) arrays.