Human civilization dynamics: why we have different civilization patterns in history

The paper addresses why human civilizations diverge into unity (exemplified by China) versus disunity (exemplified by Europe) following the Axial Age. It situates the question in theories of state formation and civilizational development—circumscription, hydraulic hypotheses, and Axial Age similarities—highlighting the Warring States period (475–221 BC) as a decisive phase that enabled China’s subsequent long-run political unity. The authors argue that war dynamics, alliance behavior, and institutional changes during this period shaped civilizational trajectories. The purpose is to use a dynamic, non-linear agent-based model to recreate this historical process, identify key mechanisms, and quantify the boundary conditions that lead to unity versus persistent fragmentation. The importance lies in moving beyond subjective, static narrative history toward computational reconstruction that can generalize conditions for different civilizational patterns.

The related research on the Warring States period is summarized in four strands: (a) Statistics of wars, documenting frequencies and scales (e.g., Chiang, 2005), enabling analysis of shifting war scales (Gunner, 2019); (b) Effects of wars on Chinese state formation and institutions—war-driven rationalization, Confucian-legalist state formation, bureaucratization, and the promotion of political unification (Zhao, 2004; 2015; Hui, 2005; Lewis, 1990); (c) Causes of wars including security dilemmas, preventive war for balance-of-power maintenance, resource competition, and misperceptions (Jervis, 1978; Levy, 1998; Van Evera, 2013); (d) War strategies and their evolution—from honor-focused conflicts under Zhou authority during Spring and Autumn to annexation-focused wars in the Warring States (Yu, 2019; Osinga, 2005). Traditional approaches are limited by subjectivity and static analysis, motivating the adoption of agent-based modeling to capture dynamic, non-linear processes (Turchin, 2006, 2018).

Design: An agent-based model (multi-agent system) is built to simulate 32 states at the onset of the Warring States (475 BC) through to unification by Qin (221 BC), comparing simulated outcomes with historical benchmarks. The model encodes mechanisms of war, alliances, betrayal, and state power dynamics to explore boundary conditions for achieving unity versus disunity.

Historical validation targets: (a) Dynamic number of states (32 to 1 by 221 BC); (b) Duration of the Warring States (255 years); (c) Territorial annexation dynamics matching maps from 475 BC to Qin unification; (d) Total number of wars (~360); (e) Total number of alliances (~60). Simulations are repeated 500 times per parameter set for robustness.

Agents and states: 32 state-agents initialized with normally distributed power Pwr in [0,100]. Each tick (year), Pwr increases by 10 up to a cap of 100, minus war costs; if Pwr ≤ 0, the state dies and its territory accrues to the winner. State power aggregates economic, political, military, population, and cultural factors.

War mechanism: Each state i has war propensity Fi ∈ (0,1), initialized as Pwr_i/100, and updated by a dynamic variable b (bellicosity): increases by b1 on victory, decreases by b1 on defeat. A war threshold y is applied. State i attacks j if Fi ≥ y and Pwr_i ≥ Pwr_j. War outcomes are determined by comparing comprehensive power (own plus allies) with a random impact term α. Outcomes yield costs: w (victory), b (balance), l (defeat); costs reduce Pwr accordingly.

Alliance mechanism: Each state has alliance propensity p_i ∈ [0,0.5]. Alliance offers are probabilistic; acceptance depends on relative power (bounded rationality): accept if Pwr_i ≥ 0.9·Pwr_j; otherwise reject. Alliances form dynamic networks contributing to comprehensive power in war.

Betrayal mechanism: Alliances dissolve opportunistically based on power. If γ_i·Pwr_j > Pwr_i, state i betrays the alliance (seeks stronger partners); otherwise remains loyal. The betrayal threshold γ ∈ [0,1] governs alliance reliability.

Power dynamics and lifecycle: Pwr_{t} = Pwr_{t-1} + 10 − Cost; if Pwr ≤ 0: death; else survival. Territory and resources transfer to winners upon annexation.

Outcome classification and parameter ranges: Civilization pattern is unity if the number of states declines from 32 to 1 within 600 ticks; otherwise disunity. Parameter sweeps cover bellicosity b1 ∈ [0.1,1.0], alliance propensity p ∈ [0.1,1.0], betrayal threshold γ ∈ [0,1], winner cost w ∈ [1,12], loser cost l ∈ [13,24]. For each combination, 500 simulations produce outcome distributions. Best-fit parameter sets minimize squared deviations from historical targets (final number of states, duration, wars, alliances).

Analytical tools: Decision Tree (CART) classifiers identify thresholds separating unity and disunity based on single-factor sweeps (bellicosity, alliance propensity, betrayal threshold, war costs). A Backpropagation Neural Network (BPNN) models mappings from five inputs (bellicosity x1, alliance propensity x2, betrayal threshold x3, winner cost x4, loser cost x5) to four outputs (final states y1, duration y2, alliances y3, wars y4). Dataset size: 29,616 items, K-fold cross-validation (k=10), five hidden layers with five neurons, 100 runs reported. Performance: MSE means (SDs) — y1: 0.00159 (0.000227); y2: 0.008753 (0.001268); y3: 0.000628 (0.000106); y4: 0.001523 (0.000433). Correlations between observed and predicted: 0.989 (y1), 0.968 (y2), 0.945 (y3), 0.953 (y4).

Model fit and replication: Under unity settings, simulations match historical benchmarks: states from 32 to ~1.22 (±1) in 255.33 ticks; wars ~359.98 (SD 87.95) near the recorded ~360; alliances ~60.79 (SD 20.55) near ~60; end-state map reducing to Qin.

Unity vs disunity contrasts: Unity runs end with ~1.22 states; disunity runs retain ~18.97 (±19) states after 700 ticks. Disunity exhibits fewer wars (155.74, SD 35.49) and similar alliances (55.21, SD 9.68) compared to unity.

Bellicosity threshold: With p=0.4, γ=0.8, w=7, l=14: when bellicosity β ≤ 0.1, simulations remain disunified; when β ≥ 0.2, most runs unify (e.g., β=0.2: 91.67% unity; β=0.3: 95.67%; β≥0.4: ≈96–99%). Under unity, increasing β reduces duration from 328.59 ticks (β=0.2) to 178.04 (β=1) and decreases alliances (77.1→51.91) and wars (343.10→254.34), stabilizing for β≥0.5 (~50 alliances; ~245–250 wars).

Alliance propensity threshold: Holding b1=0.2, γ=0.8, w=7, l=14, increasing p transitions from unity to disunity. Most simulations unify for p<0.3–0.4; disunity dominates for p≥0.4. Under unity, as p increases from 0 to 0.3, duration rises (169.75→507.18 ticks), alliances increase (0→72.41), and wars increase (276.89→354.83).

Betrayal threshold boundary: With b1=0.2, p=0.4, w=7, l=14, higher betrayal thresholds favor unity; boundary around γ_b≈0.8. Example unity proportions: γ_b=0.9 (83.33%), γ_b=1 (86.01%). Under unity, higher γ_b shortens duration (575.63→351.31) with minimal impact on alliance and war counts.

War cost effects: Higher war costs generally reduce duration, alliances, and wars. Winner cost w from 1→12 reduces duration (679.33→431.95), alliances (214.14→52.57), wars (969.17→194.38). Loser cost l from 13→24 reduces duration (679.33→377.33), alliances (214.14→61.72), wars (969.17→239.53). Loser cost has a stronger effect on shortening duration than winner cost.

Global boundaries map: Parameter sweeps show large regions yielding unity; disunity arises with very low bellicosity (e.g., b1≈0.0–0.1), high alliance propensity, or combinations including high winner costs in some settings. The study reports that higher winner cost beyond ~5% impedes unity.

Neural network validation: BPNN predicts outcomes from inputs with high fidelity (correlations 0.945–0.989), supporting a structured “social knowledge” representation that aligns with simulated and historical patterns.

The findings quantify conditions under which a multi-state system converges to political unity versus persistent fragmentation. Reproducing the Warring States trajectory, the model shows that moderate-to-high bellicosity accelerates annexation and unification, while a higher propensity to ally—reflecting hedging and shifting coalitions—impedes consolidation and favors disunity. Alliance reliability (high betrayal threshold) strengthens centripetal dynamics by stabilizing coalitions around stronger powers, thus enabling successful annexations and faster unification. War costs shape tempo and intensity: especially high loser costs hasten unification by discouraging prolonged conflict networks, while high winner costs can stall consolidation and, in some regimes, prevent unity. These quantified boundaries illuminate why China, during the Warring States period, crossed threshold conditions conducive to unification, while other regions (e.g., post-Roman Europe) remained in disunity regimes. The complementary BPNN analysis indicates that the parameter-outcome relationships are learnable and predictive, suggesting a transferable knowledge structure for historical learning about state consolidation processes.

The study advances a dynamic, evidence-constrained ABM that reconstructs the Warring States process and solves boundary conditions distinguishing unity from disunity. Key contributions include: (1) Identification of thresholds—bellicosity around ≥0.2 for unity; alliance propensity thresholds around 0.3–0.4 where higher values favor disunity; betrayal threshold around ≥0.8 favoring unity; and differential impacts of war costs (loser cost more strongly shortening duration). (2) Empirical validation by matching historical targets (state count trajectory, duration ~255 years, ~360 wars, ~60 alliances). (3) A generalizable framework for civilizational pattern emergence and a neural-network-based “social knowledge” representation that accurately predicts outcomes. Future research should integrate additional real-world variables (geography, spatial relations, climate), richer behavioral rules, and extend modeling from phase-specific (Spring and Autumn; Warring States) to the integrated Eastern Zhou and subsequent periods to probe continuity in learning and unification dynamics.

The model is idealized and omits potentially important non-modeled factors such as detailed geography, spatial proximity and routes, climate, and technological heterogeneity. Parameterization of alliance and betrayal rules abstracts complex motives and institutions. Unity/disunity classification relies on a 600-tick threshold and may miss nuanced outcomes. Some reported thresholds and effects may be context-dependent on fixed parameters during single-factor sweeps. The study focuses on the Warring States phase rather than the integrated Eastern Zhou as a whole; future work should incorporate broader temporal scope and more complex agent behaviors.