Variety

Variety: Injecting Humanity into the Machine

[cite_start]Standard chess engines are typically deterministic: given the same position, time control, and thread count, they will almost always play the exact same move[cite: 1080]. While efficient for analysis, this makes them predictable opponents and poor sparring partners.

[cite_start]ShashChess 41 introduces the Variety system, a stochastic mechanism integrated directly into the Quiescence Search[cite: 1082]. [cite_start]Unlike simple "MultiPV" randomization which picks suboptimal moves, Variety introduces controlled noise into the evaluation of leaf nodes, simulating human factors like "optimism" and "fighting spirit"[cite: 1083, 1084].

---

1. Technical Implementation

The logic is located within the search.cpp file. [cite_start]It modifies the static evaluation of a quiet position by adding a pseudo-random increment before returning it to the Alpha-Beta search tree[cite: 1086, 1087].

The Algorithm: Bounded Randomness

The system applies a dynamic bonus based on a Pseudo-Random Number Generator (PRNG). The formula ensures safety by being self-regulating:

[cite_start]$$Eval_{final} = Eval_{static} + \text{Random}(0, Limit - |Eval_{static}|)$$ [cite: 1090]

• Self-Regulating: If the position is exactly equal (0.00), the noise is maximal. [cite_start]As the position becomes decisive (won or lost), the noise decreases to zero[cite: 1110, 1111].
• Safety: This prevents the engine from hallucinating in clearly won or lost positions. [cite_start]It only adds "flavor" to balanced or unclear positions[cite: 1112, 1113].
• Optimism Bias: The increment is always added, never subtracted. [cite_start]This encourages the engine to be "optimistic," playing lines it might otherwise discard as too drawish[cite: 1114, 1115].

// Logic snippet from search.cpp
int maxValidIncrement = maxIncrement - std::abs(bestValue);
if (maxValidIncrement < 0) { maxValidIncrement = 0; }

// Inject randomness
int increment = static_cast<int>(rng.rand<uint64_t>() % (maxValidIncrement + 1));
bestValue += increment;

2. Configuration Modes

The system offers three levels of operation via the UCI option Variety:

A. Off (variety = 0)

• Behavior: Standard deterministic engine behavior.
• Use Case: Precise analysis and engine-vs-engine rating testing where reproducibility is required.

B. Standard (variety = 1)

• Limit: 13 centipawns (approx. 0.13 pawns).
• Effect: Adds a subtle noise layer. The engine remains objectively very strong with no measurable Elo loss, but it breaks deterministic loops.
• Use Case: Sparring partners, opening book generation, and avoiding identical games from the same starting position.

C. Psychological (variety = 2)

• Limit: 309 centipawns (approx. 3 pawns).
• Effect: Simulates a human player's "fighting spirit". In an objectively equal position (0.00), the engine might "believe" it has a significant advantage due to the noise.

Tal-Like Bluffing:

This prompts the engine to play aggressively rather than steering for a draw. It may enter complications that are objectively unsound but practically difficult to refute.

• Use Case: Playing against humans to avoid "boring" computer draws or training to defend against speculative attacks.

3. Integration with Shashin Theory

The Variety option perfectly complements the Shashin philosophy:

Chaos Zones:

The Psychological mode effectively forces the engine into Chaos Zones (where $\Delta K \neq 0$).

Anti-Draw Technology:

By artificially inflating evaluations, it mimics the High Tal desire to complicate the game, rejecting the safety of Capablanca (stable) evaluations.

Goal:

"Fewer Draws, More Fun." Even if the engine calculates a line to be a draw, the Psychological noise pushes it to continue playing for a win, simulating human ambition.