🤖 AI

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

This paper demonstrates that while Hebbian plasticity initially reduces behavioral variance in neuroevolved chess agents, it ultimately expands diversity over evolutionary time through three distinct regimes, challenging the notion that learning always buffers against environmental noise and suggesting that self-play systems may inadvertently suppress the heterogeneity necessary for personality emergence.

Original authors: Diego Armando Resendez Prado

Published 2026-04-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Diego Armando Resendez Prado

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are training a team of chess players. But instead of teaching them the rules of chess, you give them a crystal ball (the "Cartridge") that predicts what moves are good. The crystal ball is decent, but not perfect.

Your goal isn't just to make the team win; you want to see if they can develop personalities. Do they all play the same way, or do some become aggressive attackers while others become cautious defenders?

This paper explores a fascinating question: Does the ability to learn during a game (plasticity) help a population become more diverse, or does it make everyone the same?

Here is the story of what happened, explained through simple analogies.

1. The Setup: The Brain and the Crystal Ball

The researchers built a "Brain" made of eight tiny neural networks (like a mini-cognitive team: Perception, Memory, Emotion, etc.). This Brain sits on top of the Crystal Ball.

The Crystal Ball (Cartridge): It looks at the board and says, "Move A is 60% likely to be good."
The Brain: It listens to the Crystal Ball but can tweak the decision. It can say, "Actually, I like Move B better because I feel lucky today," or "No, let's stick to Move A."

The Brain has two ways of getting smarter:

Evolution (The Long Game): Every 50 games, the best players get to have "babies" (mutations) to pass on their traits. This happens over many generations.
Hebbian Learning (The Short Game): During a single game, the Brain learns from its mistakes immediately. If a move feels right and wins, it strengthens that connection. If it loses, it weakens it.

2. The Big Surprise: The "Personality" Crossover

The researchers ran two groups of teams for 50 generations:

Group A (Hebbian ON): The brains could learn during the game.
Group B (Hebbian OFF): The brains could only learn through evolution (no in-game learning).

The Prediction:
Old theories said Group A should be more uniform. Because they learn quickly from the environment, they should all figure out the "perfect" way to play and converge on the same strategy. Group B, being slower and relying on luck, should be more chaotic and diverse.

The Reality (The Twist):

Early on (Generations 1–33): The prediction was right! Group A (learning brains) was very similar. They all learned the same basic lessons quickly and played alike. Group B was messy and diverse.
The Crossover (Generation 34): Suddenly, Group A exploded with diversity. They became more different from each other than Group B ever was.
Why? The "learning during the game" acted like a magnifying glass.
- Imagine two seeds (two different starting brains) have a tiny, random difference in how they "see" the board.
- In Group B, this tiny difference is lost in the noise of evolution.
- In Group A, the brain uses its "Imagination" (a mental simulation of the next move) to test these tiny differences. The learning process reinforces these tiny quirks. One seed starts to love Knights; another starts to hate them. Over time, these small quirks get amplified into completely different playing styles.

3. The Three Regimes (The Three Ways the Game Plays Out)

The paper found that the outcome depends entirely on who the opponent is.

Regime 1: The Explorer (Different Opponent + Learning ON)

Scenario: Your brain plays against a different type of engine (Maia2).
Result: Personality Emerges.
- Some brains became "Diplomats": They played safe, agreed with the crystal ball, and played short, quick games.
- Others became "Fighters": They played risky, disagreed with the crystal ball, and played long, chaotic games.
- The Lesson: When the environment is complex and different from your internal model, learning helps you specialize. You don't just find one answer; you find many unique ways to win.

Regime 2: The Lottery (Different Opponent + Learning OFF)

Scenario: Your brain plays against a different engine, but it cannot learn during the game.
Result: The "Jackpot" Effect.
- Most brains stayed mediocre.
- But occasionally, one brain got a "lucky mutation" that just happened to work perfectly. Once it won, it was cloned forever (Elitism).
- The population didn't evolve diverse strategies; it just got stuck on a few lucky winners. It's like a lottery where everyone buys a ticket, but only one person wins big, and everyone else copies them.

Regime 3: The Mirror (Same Opponent + Learning ON)

Scenario: Your brain plays against a copy of itself (or the same engine).
Result: Self-Erasure.
- The brains realized: "If I change my strategy, I'm just playing against myself, and I'll lose."
- They stopped trying to be unique. They became transparent. They just did exactly what the crystal ball said, 100% of the time.
- The Lesson: If you only play against yourself (Self-Play), you lose your personality. You become a mirror, reflecting only the "average" move. This suggests that AI systems that only train against themselves (like AlphaZero) might be suppressing their own potential for diverse, creative strategies.

4. The Takeaway: Why This Matters

"Personality Requires Struggle."
The paper argues that to develop a unique "personality" (a distinct way of solving problems), you need:

A weak starting point: You need a system that isn't perfect yet (the Crystal Ball wasn't perfect).
A different opponent: You need to face something that doesn't think exactly like you.
The ability to learn: You need to be able to adapt while you are doing the task.

The Big Picture:
In the past, we thought learning made everyone the same (convergence). This paper shows that in a complex, competitive world, learning actually helps a population explode into diversity. It allows different individuals to find different "niches" and develop unique strengths.

However, if you remove the "otherness" (by playing against yourself), that diversity vanishes, and everyone becomes a boring, identical copy.

In short: To build a world of unique, creative agents, don't just let them play against themselves. Let them learn, let them struggle, and let them face opponents who think differently than they do. That is where "personality" is born.

1. Problem Statement

The paper investigates a fundamental question in evolutionary biology and artificial intelligence: Does lifetime learning (plasticity) expand or collapse behavioral diversity over evolutionary time?

Prior Theory: Existing theories (e.g., Paenke et al.) suggest that plasticity reduces phenotypic variance by buffering organisms against environmental noise, accelerating convergence toward a single optimum.
The Gap: Most prior computational demonstrations of the Baldwin Effect (where learning ability influences evolution) rely on "toy domains" (binary strings, simple functions) with single-basin fitness landscapes. It is unknown how plasticity interacts with evolution in complex, competitive domains with multi-modal fitness landscapes.
Hypothesis: In a competitive environment like chess, plasticity might not just buffer against noise but could open new regions of the fitness landscape, leading to increased behavioral diversity (personality) over time.

2. Methodology

2.1 The AILED-Brain Architecture

The authors introduce AILED-Brain, a modular cognitive architecture evolved via NEAT (NeuroEvolution of Augmenting Topologies). It sits on top of a "cartridge" (a pre-trained move predictor) and reshapes its output.

Modules: Eight NEAT-evolved neural modules: Perception, Memory, Affect, Attention, Dynamics, Personality, Integration, and Learning.
Signal Chain: The modules output parameters for a "desirability-domain" signal chain. Instead of directly predicting moves, the brain modifies the desirability of moves based on:
- Imagination: A 1-ply subjective lookahead where the brain evaluates candidate moves through its own evolved Perception module.
- Piece Preferences: Weights biasing moves involving specific piece types.
Timescales:
1. Genetic Evolution (NEAT): Occurs across generations (population size 20). Selects genomes based on fitness.
2. Lifetime Learning (Hebbian Plasticity): Occurs within a single game. Connection weights are adjusted based on correlated activations and game outcomes ( $\Delta w = \eta \cdot a_{pre} \cdot (r - b)$ ).

2.2 Experimental Setup

Cartridge (Base Predictor): AILED Engine v26.3.0, a 43M-parameter transformer trained on human games. Crucially, it exposes Win/Draw/Loss (WDL) estimates and piece-level attention (perceptual signals), not just move probabilities.
Opponents:
- Heterogeneous: Maia2 (1100 Elo) – a different engine with different error patterns.
- Homogeneous (Mirror): Maia2 vs. Maia2 (various Elo levels) – same engine, same error patterns.
Conditions:
- Hebbian ON: Agents learn within the game.
- Hebbian OFF: Agents rely solely on genetic mutation (no lifetime learning).
- Ablations: Removing "Imagination" or varying fitness weights.
Metrics: Fitness (Win rate + Agreement + Calibration), Agreement (match with cartridge), and Intraclass Correlation Coefficient (ICC) to measure reproducible behavioral divergence ("personality").

3. Key Contributions

Variance Crossover Phenomenon: The paper demonstrates that plasticity's effect on variance is time-dependent. Initially, Hebbian learning compresses diversity (lower variance than non-plastic agents), but after a crossover point (Generation 34), it expands diversity, surpassing non-plastic agents.
Three Distinct Regimes: The interaction between the brain, the cartridge, and the opponent creates three qualitatively different evolutionary dynamics:
- Exploration (Hebbian ON + Heterogeneous Opponent): High diversity, continuous improvement, and distinct "personalities."
- Lottery (Hebbian OFF + Heterogeneous Opponent): Low diversity, characterized by "lucky" mutations followed by elitism lock-in.
- Transparent (Same-Model Opponent): The brain "self-erases," converging to 100% agreement with the cartridge regardless of skill gap.
Structured Behavioral Divergence: Evolved agents develop reproducible, interpretable "personalities" (e.g., distinct opening repertoires, piece preferences, and game lengths) with high ICC (>0.8).
Falsifiable Prediction on Self-Play: The authors predict that self-play systems (where agent A plays agent A) systematically suppress behavioral diversity because the "mismatch" required for personality to emerge is absent.

4. Key Results

4.1 The Variance Crossover (Exploration Regime)

Early Evolution (Gen 1–33): Hebbian ON agents have lower cross-seed variance than OFF agents. Plasticity buffers them, causing them to learn similar early lessons.
Late Evolution (Gen 34+): Hebbian ON agents surpass OFF agents in variance. The variance ratio increases monotonically ( $\rho = 0.91, p < 10^{-6}$ ).
Mechanism: Evolved differences in the Perception module are amplified by the Imagination loop. Since different seeds have different Perception weights, the same board position "feels" different to each brain, leading to divergent strategies.
Outcome: Seeds develop distinct strategies (e.g., "Diplomat" vs. "Fighter").
- Diplomat: High agreement with cartridge, short games, knight-focused.
- Fighter: Low agreement (anti-agreement), long games, queen-focused.
- Disagreement: Agents select different moves on the same board 62% of the time.

4.2 The Lottery Regime (Hebbian OFF)

Without lifetime learning, evolution relies on discrete "lucky" jumps.
Lock-in: Once a decent genome is found, elitism freezes the population. 8/10 OFF seeds converged to exactly 50% win rate, whereas ON seeds showed a wide spread (17%–65%).
Stagnation: The late-stage fitness slope for OFF agents is near zero, indicating a lack of incremental improvement.

4.3 The Transparent Regime (Same-Model Opponents)

When the cartridge and opponent are the same model (e.g., Maia2 vs. Maia2), the brain learns to do nothing.
Self-Erasure: Agreement converges to 100% within 7–13 generations. The brain realizes that any modulation hurts performance against a twin.
Control: When the opponent is a different engine (Stockfish) at the same Elo, the brain retains diversity (~30% agreement), proving that self-erasure is an artifact of identical error patterns, not just high skill.

4.4 Ablation Studies

Imagination Removal: Removing the 1-ply lookahead collapsed the variance crossover. Without imagination, different Perception weights did not translate into behavioral differences (variance dropped to near zero).
Hebbian Removal on Evolved Genomes: Evolved "ON" genomes performed significantly worse without Hebbian learning, confirming that the genome encodes learnability (Baldwin Effect) rather than just static competence.

5. Significance and Implications

Reversal of Prior Theory: The paper challenges the view that plasticity always reduces variance. It shows that in complex, competitive landscapes, plasticity can act as a diversity amplifier over long timescales.
Personality Requires Struggle: "Personality" (systematic behavioral divergence) only emerges when there is a mismatch between the agent's internal model (cartridge) and the external environment (opponent). If the environment is perfectly predictable or identical to the self (self-play), diversity collapses.
Critique of Self-Play: The findings suggest that AlphaZero-style self-play systems may inadvertently suppress the emergence of diverse play styles or "personalities" because they eliminate the heterogeneity required for the Baldwin Effect to drive divergence.
Expressiveness > Accuracy: A weak predictor (21.3% accuracy) paired with an evolved cognitive layer outperformed strong predictors with simple temperature tuning. Rich perceptual outputs (WDL, attention) are more critical for cognitive evolution than raw prediction accuracy.
Baldwin Effect Validation: The study provides empirical evidence of the Baldwin Effect in a multi-module cognitive architecture, showing how learned behaviors (Hebbian) reshape the fitness landscape to select for genetic variants that approximate those behaviors.

Conclusion

The paper concludes that plasticity's effect on behavioral variance is non-monotonic: it compresses diversity initially but expands it later through the interaction of evolved perception and imagination. This process requires a heterogeneous environment; in homogeneous (self-play) environments, the system converges to a "transparent" state where the cognitive layer erases itself. This offers a new framework for understanding the emergence of personality in AI and the limitations of self-play training regimes.