Neuroevolution Arena: Nested Ecological Evaluation of Update-and-Inheritance Regimes across Neural Architectures

Neuroevolution Arena presents a GPU-accelerated artificial life platform in which up to 10,000 cells, each equipped with an independent neural network, compete on a toroidal grid. Unlike standard Neural Cellular Automata (NCA) where all cells share a single learned update rule, this system supports heterogeneous populations: cells may differ in network architecture and in the update-and-inheritance rules applied to W1 and W2. Three implementation-specific update-and-inheritance regimes—EvoEvo, EvoRL, and RLRL—are crossed with two neural architectures for 50,000 generations in three independent training runs per condition. EvoEvo applies evolutionary crossover and mutation to both layer groups, EvoRL combines W1 evolution with lifetime RL on W2, and RLRL applies lifetime RL to both while retaining reproductive inheritance and selection. One saved elite-controller artifact from each of the 18 runs enters an aligned-run frozen-evaluation design comprising 198 computational jobs: 135 pairwise competitions, 9 six-way melees, and 54 survival jobs. Pairwise effects average three seed-defined ecological contexts—two cooperation-permitting and one attack-permitting—within each aligned training-run block; the independent level remains n = 3 runs per condition.

In training results, both RL-enabled regimes had higher recorded training fitness than EvoEvo in the MLP4 subset: EvoRL reached 335K and RLRL 299K, compared with 150K for EvoEvo. EvoEvo maintained the largest population (6,008) but the shortest lifetime (29.1 steps), whereas EvoRL and RLRL maintained smaller populations (5,071 and 4,905) with longer lifetimes (67.1 and 62.8). The training analysis also showed a crossover interaction: EvoRL had a positive interaction term with MLP4 (+21,504 relative to the additive prediction) and a negative term with Wide128 (-21,504), while RLRL showed the opposite pattern (+16,209 for Wide128 and-16,209 for MLP4).

In formal frozen competition results, EvoEvo×MLP4 had the largest pairwise run-block count under this protocol, winning 11 of its 15 opponent-by-aligned-run blocks and holding a 2-of-3 or 3-of-3 run-block majority against each of its five opponents. EvoRL×MLP4, RLRL×MLP4, and RLRL×Wide128 each won 9 of 15 blocks; EvoRL×Wide128 won 5, and EvoEvo×Wide128 won 2. With n = 3 aligned training-run blocks, the descriptive majority ordering reverses across architectures. Within MLP4, EvoEvo beat EvoRL by 2 blocks to 1, EvoEvo beat RLRL 2–1, and EvoRL beat RLRL 2–1, giving EvoEvo > EvoRL > RLRL. Within Wide128, RLRL beat EvoRL 2–1 and EvoEvo 3–0, while EvoRL beat EvoEvo 2–1, giving RLRL > EvoRL > EvoEvo. In addition, MLP4 beat Wide128 by 2–1 within each regime. Only 4 of the 15 condition pairs retained the same effect direction across all three aligned run blocks; the other 11 reversed direction in one block. At the nested-context level, only 17 of 45 pair-by-run blocks produced the same winner under all three contexts.

The nine six-way jobs were even more diverse in their winners: EvoEvo×MLP4 won 4 of 9, RLRL×MLP4 won 2, and EvoRL×MLP4, EvoRL×Wide128, and RLRL×Wide128 each won 1; EvoEvo×Wide128 won none. Eight of nine final states contained only one surviving origin.

The replicated frozen survival results produced a complete floor effect. None of the 54 jobs reached the 5,000-cell threshold after any 97% attempt, so every condition had a maximum recovered ratio of 0. No job advanced to the planned 98% or 99% stages. Fifty jobs remained non-extinct after their attempted 97% shocks, while four became extinct. A secondary, descriptive endpoint—the final population after attempted shocks—was highest for EvoRL×MLP4 (mean of training-run means 2,430 ± 410), followed by RLRL×Wide128 (2,194 ± 383), EvoEvo×MLP4 (2,121 ± 226), RLRL×MLP4 (2,036 ± 182), EvoEvo×Wide128 (1,697 ± 332), and EvoRL×Wide128 (1,619 ± 736).

The study contributes a nested protocol that separates training-run artifacts from evaluation contexts and exposes, rather than conceals, their different sources of variation. The central empirical result is that the apparent ordering of update regimes depends on architecture, and the limited directional agreement across aligned artifacts prevents any condition from being treated as a universal competitive leader. Across evaluation dimensions, no single condition dominates: training fitness is highest for RLRL×Wide128, EvoEvo×MLP4 has the largest aligned-run pairwise count, and the nine six-way jobs distribute wins across five conditions. The strict survival endpoint yields no winner because all 54 scores are zero. The results support a bounded claim: ecological outcomes in this system depend jointly on controller configuration, saved training-run artifact, opponent set, and evaluation context; they do not identify a universally superior regime, a survival leader, or a demonstrated Baldwin Effect.

Improvements to AI systems:

1. Architecture-Dependent Learning Regime Selection: Implement a meta-controller that dynamically selects between evolutionary (W1/W2 crossover+mutation), hybrid (evolution + lifetime RL), or pure lifetime RL update rules based on the neural network architecture being trained. For MLP4-type small networks, prioritize evolutionary updates; for Wide128-type wide networks, prioritize RL updates. This exploits the observed crossover interaction (EvoRL positive for MLP4, RLRL positive for Wide128) to maximize training fitness.

2. Heterogeneous Population Training with Inheritance Rules: Replace single-rule NCA training with a heterogeneous population where each cell can have different network architectures and different update/inheritance rules for separate weight groups (W1, W2). The improved system can evolve mixed populations that balance population size (e.g., 5,000–6,000 cells) and lifetime (29–67 steps) adaptively, avoiding the trade-off seen in EvoEvo (large population, short lifetime) versus RL regimes (smaller population, longer lifetime).

3. Context-Aware Evaluation and Robustness Scoring: Build an evaluation pipeline that tests trained agents across multiple ecological contexts (cooperation-permitting vs. attack-permitting) and multiple independent training-run artifacts, not just a single run. The improved system can compute a directional agreement score (e.g., fraction of aligned runs where the same winner emerges) and reject any agent that reverses its competitive ordering across contexts, preventing deployment of brittle solutions.

4. Survival Shock Resilience via Multi-Threshold Training: Train agents with a survival objective that includes incremental shock recovery (e.g., 97%, 98%, 99% cell removal) rather than a single binary threshold. Since all 54 jobs failed at 97% recovery, the improved system can use a curriculum of increasing shock severity, with a secondary metric (final population after shock) as a continuous reward signal, enabling graded improvement instead of all-or-nothing failure.

5. Artifact-Aware Model Selection: When deploying a trained controller, do not select based solely on training fitness or pairwise wins. Instead, use a multi-dimensional leaderboard that includes: (a) pairwise win count across aligned runs, (b) six-way melee survival, (c) post-shock population, and (d) consistency across contexts. The improved system can avoid choosing a false leader like EvoEvo×MLP4 (which wins pairwise but fails survival) or RLRL×Wide128 (which wins training fitness but not pairwise).

6. Baldwin Effect Verification Module: Add a diagnostic that explicitly tests whether lifetime learning (RL on W2) leads to heritable improvements in W1 evolution across generations. Since no Baldwin Effect was demonstrated, the improved system can monitor the correlation between W2 learned weights and W1 evolutionary fitness over time, and if absent, switch to pure evolution or pure RL to avoid wasted computation.

What the improved AI system can do:

• Automatically select the optimal training paradigm (evolution, hybrid, or RL) based on the target network architecture, improving training fitness by up to 2.2× (e.g., from 150K to 335K) for small networks.

• Evolve heterogeneous populations that self-balance population size and individual lifetime, leading to more stable long-term training without premature extinction.

• Produce agents that are robust across ecological contexts and training-run artifacts, with a quantified consistency score, reducing the risk of overfitting to a single evaluation scenario.

• Recover from severe population shocks (e.g., 97% cell removal) by learning graded survival strategies, instead of failing completely.

• Provide a multi-criteria evaluation report that prevents selection of brittle champions, improving real-world deployment reliability in competitive multi-agent environments.

• Detect and report the presence or absence of the Baldwin Effect, allowing researchers to decide whether to invest in lifetime learning for their specific problem domain.