cs.AIApr 4, 2026

Personality Requires Struggle: Three Regimes of the Baldwin Effect in Neuroevolved Chess Agents

Authors: Diego Armando Resendez Prado

Organizations: Independent Researcher

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Can lifetime learning expand behavioral diversity over evolutionary time, rather than collapsing it? Prior theory predicts that plasticity reduces variance by buffering organisms against environmental noise. We test this in a competitive domain: chess agents with eight NEAT-evolved neural modules, Hebbian within-game plasticity, and a desirability-domain signal chain with imagination. Across 10seeds per Hebbian condition, a variance crossover emerges: Hebbian ON starts with lower cross-seed variance than OFF, then surpasses it at generation34. The crossover trend is monotonic (\r{ho} = 0.91, p < 10^{-6): plasticity's effect on behavioral variance reverses over evolutionary time, initially compressing diversity (consistent with prior predictions) then expanding it as evolved Perception differences are amplified through imagination -- a feedback loop that mutation alone cannot sustain. The result is structured behavioral divergence: evolved agents select different moves on the same positions (62% disagreement), develop distinct opening repertoires, piece preferences, and game lengths. These are not different sampling policies -- they are reproducible behavioral signatures (ICC > 0.8) with interpretable signal chain configurations. Three regimes appear depending on opponent type: exploration (Hebbian ON, heterogeneous opponent), lottery (Hebbian OFF, elitism lock-in), and transparent (same-model opponent, brain self-erasure). The transparent regime generates a falsifiable prediction: self-play systems may systematically suppress behavioral diversity by eliminating the heterogeneity that personality requires. \textbf{Keywords: Baldwin Effect, neuroevolution, NEAT, Hebbian learning, chess, cognitive architecture, personality emergence, imagination

Figures & tables

Figure 1: Pool architecture. Phase 1: five modules read independently from a shared signal pool (board sensors, game context, cartridge WDL, distribution shape). Phase 2: Personality reads all Phase 1 outputs. Phase 3: Integration produces signal chain parameters. The signal chain reshapes the cartridge’s move distribution; imagination evaluates top candidates through 1-ply lookahead using Perception (dashed). Each module is a NEAT network.
PhModuleDimsRole
1Perception20 →\to 8Board feature compression
1Memory8 →\to 4Game trajectory context
1Affect8 →\to 5Emotional response (WDL)
1Attention16 →\to 4Selective focus (pieces)
1Dynamics6 →\to 2Confidence calibration
2Personality27 →\to 8Preference shaping
Table 1: Cognitive modules and execution phases.
Figure 2: Neuroscience inspiration for the AILED-Brain architecture. Left: human brain regions. Right: corresponding AILED-Brain modules. Dashed gray arrows show the functional mapping; color coding indicates analogous roles. Both systems feature parallel sensory processing, preference integration, and action selection. The key parallel: imagination reuses Perception (blue dashed) the same way mental imagery reuses visual cortex. Bottom: both operate on two timescales—within-lifetime plasticity and cross-generation evolution.
ExperimentCartridgeOpponentHebbGens
Set 1: Hebbian ON vs OFF (10 seeds each)
v26-ONAILED Engine v26.3.0Maia2 1100ON50
v26-OFFAILED Engine v26.3.0Maia2 1100OFF50
Set 2: Same-model (mirror) matchups
DominantMaia2 1400Maia2 1100ON100
Match-1400Maia2 1400Maia2 1400ON100
Table 2: Experimental conditions. All: population 20, desirability-domain signal chain, multi-objective fitness unless noted.
MetricHebbian ONHebbian OFF
Seeds ( nn )1010
Mean fitness0.471±0.0470.471\pm 0.0470.485±0.0280.485\pm 0.028
Mean agreement40.4%±7.4%40.4\%\pm 7.4\%41.8%±4.8%41.8\%\pm 4.8\%
Agreement range31.7%–54.1%33.9%–48.0%
Mean win rate0.42±0.160.42\pm 0.160.49±0.120.49\pm 0.12
Seeds at win == 0.501 / 108 / 10
Table 3: Hebbian ON vs OFF with AILED Engine v26.3.0 cartridge against Maia2 at 1100 Elo. 50 generations, population 20, 20 games/eval. 10 seeds per condition.
Figure 3: Hebbian ON vs OFF trajectories (mean across 10 seeds each) with AILED Engine v26.3.0 against Maia2 at 1100 Elo. Left: mean best fitness. OFF rises faster but plateaus by gen 20; ON climbs steadily through gen 50. Right: mean agreement. Both approach ∼\sim 40–42%, but ON’s late-stage slope is 7.6×7.6\times larger than OFF’s. The difference is in dynamics, not endpoint.
ExperimentOpponent typeAgreeGens
1400 vs 1100Same (dominant)50% ∗100
1400 vs 1400Same (mirror)100% at g13100
1600 vs 1600Same (mirror)100% at g750
1400 vs 1600Same (underdog)100% at g820
1600 vs SF 1600Different (3 seeds)30.7% ±\pm 1.3%20
∗ Flat at 50% throughout; brain contributes nothing.
Table 4: Same-model and heterogeneous opponent experiments (all Hebbian ON).
Figure 4: Agreement trajectories across the three regimes. (a) Expressive cartridge (AILED Engine v26.3.0 vs Maia2-1100, mean of 10 seeds per condition): both conditions approach ∼\sim 40% agreement, but ON is still climbing while OFF has flatlined. The difference is in dynamics, not endpoint. (b) Dominant cartridge (Maia2 1400 vs 1100, first 25 of 100 generations shown): agreement flat at 50% throughout—the brain is passive. (c) Mirror matches: agreement converges to 100% (self-erasure). Maia2-1600 reaches transparency at gen 7, faster than 1400 at gen 13.
  vs 1100vs 1400vs 1600
BrainHebbWinAgrWinAgrWinAgr
DiplomatOFF25%44%32%46%25%48%
 ON41%52%44%52%37%48%
FighterOFF34%20%25%25%17%20%
 ON25%21%4%19%5%22%
Table 5: Hebbian ablation on two evolved ON genomes vs Maia2 at three Elo levels (100 games each). Same genome; only lifetime learning differs.
Figure 5: Parameter evolution of the diplomat (seed 2) and fighter (seed 9) across generations 10–50. Top: saturation ceiling diverges from similar starting values to opposite extremes (0.91 vs 0.22). Bottom: temperature modifier diverges from near-zero to opposite signs ( +0.26+0.26 vs −0.44-0.44 ). Divergence is learned progressively, not present at initialization.
More in this topic

Neuroevolution

Neuroevolution uses evolutionary algorithms to optimize neural-network weights, architectures, and learning rules.

See all 36 papers

Explore similar work

CardsList
Apr 3, 2026

Activity-Dependent Plasticity in Morphogenetically-Grown Recurrent Networks

Developmental approaches to neural architecture search grow functional networks from compact genomes through self-organisation, but the resulting networks operate with fixed post-growth weights. We characterise Hebbian and anti-Hebbian plasticity across 50,000 morphogenetically grown recurrent controllers (5M+ configurations on CartPole and Acrobot), then test whether co-evolutionary experiments -- where plasticity parameters are encoded in the genome and evolved alongside the developmental architecture -- recover these patterns independently. Our characterisation reveals that (1) anti-Hebbian plasticity significantly outperforms Hebbian for competent networks (Cohen's d = 0.53-0.64), (2) regret (fraction of oracle improvement lost under the best fixed setting) reaches 52-100%, and (3) plasticity's role shifts from fine-tuning to genuine adaptation under non-stationarity. Co-evolution independently discovers these patterns: on CartPole, 70% of runs evolve anti-Hebbian plasticity (p = 0.043); on Acrobot, evolution finds near-zero eta with mixed signs -- exactly matching the characterisation. A random-RNN control shows that anti-Hebbian dominance is generic to small recurrent networks, but the degree of topology-dependence is developmental-specific: regret is 2-6x higher for morphogenetically grown networks than for random graphs with matched topology statistics.

Sergii Medvid, Andrii Valenia, Mykola Glybovets

National University of Kyiv-Mohyla Academy, Kyiv, Ukraine

Recurrent Neural NetworksNeuroevolutionHebbian LearningNeural Processes
AskPDF
Mar 5, 2026

Ailed: A Psyche-Driven Chess Engine with Dynamic Emotional Modulation

Chess engines passed human strength years ago, but they still don't play like humans. A grandmaster under clock pressure blunders in ways a club player on a hot streak never would. Conventional engines capture none of this. This paper proposes a personality x psyche decomposition to produce behavioral variability in chess play, drawing on patterns observed in human games. Personality is static -- a preset that pins down the engine's character. Psyche is dynamic -- a bounded scalar ψ_t \in [-100, +100], recomputed from five positional factors after every move. These two components feed into an audio-inspired signal chain (noise gate, compressor/expander, five-band equalizer, saturation limiter) that reshapes move probability distributions on the fly. The chain doesn't care what engine sits behind it: any system that outputs move probabilities will do. It needs no search and carries no state beyond ψ_t. I test the framework across 12,414 games against Maia2-1100, feeding it two probability sources that differ by ~2,800x in training data. Both show the same monotonic gradient in top-move agreement (~20-25 pp spread from stress to overconfidence), which tells us the behavioral variation comes from the signal chain, not from the model underneath. When the psyche runs overconfident, the chain mostly gets out of the way (66% agreement with vanilla Maia2). Under stress, the competitive score falls from 50.8% to 30.1%. The patterns are reminiscent of tilt and overconfidence as described in human play, but I should be upfront: this study includes no human-subject validation.

Diego Armando Resendez Prado

Independent Researcher

Game-Playing AgentsHuman Behavior SimulationChessAffective Computing
AskPDF
Sep 29, 2026

Harness Evolution as Learning: Approximation, Generalization, and Optimization Limits of Self-Improving Personal Agents

As the capabilities of large language models (LLMs) continue to advance, increasing attention is turning to how to translate their abilities into useful behavior. Personal agents bring this question into everyday settings, where models are expected to serve individual users and continually adapt to their preferences. With the underlying model held fixed, such adaptation relies on harness engineering: designing and evolving the surrounding layer that manages context, memory, tools, and execution. Despite rapid progress, the factors governing effective harness evolution remain insufficiently understood. To narrow this gap, we investigate three central questions concerning harness architecture, harness scale, and self-evolution algorithms through complementary empirical and theoretical analyses. Empirically, we introduce a preference-oriented benchmark and systematically characterize the capabilities and limitations of personal agents associated with these three dimensions. Theoretically, we formulate harness evolution as a learning problem and explain these phenomena through approximation, generalization, and optimization errors. Analyses of reachable policies, capacity under finite interaction evidence, and biased update dynamics provide theoretical accounts of the observed phenomena. Together, these results offer a unified perspective on the limits of personalization through harness evolution and inform future harness design.

Zeyu Gan, Zixuan Gong, Yong Liu

Gaoling School of Artificial Intelligence Renmin University of China Beijing, China

Agent Harness OptimizationLLM Agent Self-ImprovementPersonalized Language Models
AskPDF