DNA Evolution Performance DARWIN ENGINE v2.0

GENESIS: Genetic Programming DNA Evolution

GENESIS evolves expression trees representing mathematical formulas. Each tree node is either a binary operator (add, sub, mul, div, max, min, gt, lt), a unary operator (neg, abs, sin, cos, tanh, log, sqrt, clip), an input feature (26 available: OHLCV, RSI, EMA, BB, MACD, ATR, OBV, VWAP), or a random constant.

The buy and sell signals are separate trees. Crossover swaps random subtrees between parents. Mutation replaces a random subtree with a freshly grown random tree. Adaptive mutation increases the rate from 0.15 to 0.40 when evolution stagnates for 3+ generations.

Fitness = weighted combination of win_rate, sharpe_ratio, profit_factor, max_drawdown, and trade_count. Hall of Fame seeding persists winners in SQLite across runs for cumulative improvement.

ATLAS: MAP-Elites Quality-Diversity

Based on Mouret & Clune (2015), MAP-Elites maintains a 5-dimensional archive grid (5x5x3x3x3 = 675 cells) where each cell holds the best strategy for a specific behavioral niche. Dimensions: trade frequency, risk profile, direction bias, regime preference, and formula complexity.

Instead of converging to one optimum, ATLAS illuminates the entire behavior space, discovering diverse strategy types: scalpers, swing traders, conservative, aggressive, mean-reversion, trend-following, etc. The QD Score (sum of all cell fitnesses) measures quality AND diversity simultaneously.

NEXUS: Audit Ensemble Meta-Evolution

NEXUS evolves a weight vector of length 40+ (one per audit source system). The genome is in logit space; fitness evaluation applies softmax normalization to get positive weights summing to 1. Fitness combines: (1) weighted signal strength per symbol (direction x confidence x historical WR), (2) diversity reward (1 - std(weights)), and (3) entropy bonus (-sum(w * log(w))).

This is unique because it doesn't need market data directly — it reads existing pick files from all 40+ trading systems and evolves the optimal way to combine their opinions.

LEGION: Ensemble Coevolution

LEGION evolves teams of 3-8 strategies that vote together. The DNA encodes: member selection, voting weights, consensus mechanism (majority/weighted/unanimous/cascade/bayesian), veto powers, and participation thresholds. Fitness is evaluated at the ensemble level — the team's collective trading performance, not any individual member's.

Crossover swaps members between parent ensembles. Mutation can: perturb weights, add/remove members, switch consensus mechanism, or adjust thresholds. A mediocre strategy might be valuable if it provides unique perspective to the team.

CORTEX: NEAT Neural Topology Evolution

Based on Stanley & Miikkulainen (2002), CORTEX evolves neural network topologies — both structure and weights simultaneously. Networks start minimal (inputs directly connected to outputs) and complexify over generations by adding nodes and connections through structural mutations.

Innovation numbers (historical markings) enable meaningful crossover between networks with different topologies. Species-based evolution protects new structural innovations from premature elimination, giving novel topologies time to optimize their weights. Each pick includes node count, connection complexity, species ID, and generation.

ADAPT: Self-Adaptive Hyperparameter DNA Evolution

ADAPT evolves 20 strategy parameters (TP/SL multipliers, position sizing, regime sensitivity, indicator periods, etc.) with a unique twist: each parameter carries its own evolving mutation rate. Parameters that strongly affect fitness develop higher mutation rates (more exploration), while stable parameters converge (lower mutation rates).

Epigenetic importance tracking measures each parameter's sensitivity to fitness changes. High-importance parameters (like stop-loss multiplier) get fine-tuned more carefully, while low-importance ones are explored broadly. This self-adaptive mechanism outperforms fixed mutation rates in complex parameter spaces.