Abstract
Natural selection drives species to develop brains, with sizes that increase with the complexity of the tasks to be tackled. Our goal is to investigate the balance between the metabolic costs of larger brains compared to the advantage they provide in solving general and combinatorial problems. Defining advantage as the performance relative to competitors, a two-player game based on the knapsack problem is used. Within this framework, two opponents compete over shared resources, with the goal of collecting more resources than the opponent. Neural nets of varying sizes are trained using a variant of the AlphaGo Zero algorithm. A surprisingly simple relation, $N_A/(N_A+N_B)$, is found for the relative win rate of a net with $N_A$ neurons against one with $N_B$. Success increases linearly with investments in additional resources when the networks sizes are very different, i.e. when $N_A \ll N_B$, with returns diminishing when both networks become comparable in size.
Introduction
In the study of evolutionary biology and artificial intelligence, a fundamental question persists: how does computational capacity translate into competitive advantage? In biological systems, "computation" (brains) is metabolically expensive, requiring evolution to balance the benefits of increased intelligence against energy constraints. While it is well-established that larger neural networks generally perform better, the functional scaling of success in adversarial, resource-constrained environments remains poorly understood.
This paper addresses this gap by examining a competitive version of the knapsack problem—a classic NP-hard combinatorial challenge. Unlike the standard solo version, the competitive variant simulates "survival of the fittest" dynamics where agents must outperform rivals to secure shared resources.
Methods
The researchers implemented a two-player extension of the 0-1 knapsack problem. Two agents take turns selecting items from a common pool of 16 items. Item weights and values were weakly correlated (weights $w_i \in [0, 1]$, values $v_i \in w_i \pm 0.1$) to ensure the problem instances were sufficiently difficult. The game concludes when neither player can add more items without exceeding their capacity ($W = n/4$).
Agents were modeled using fully connected feed-forward neural networks with two hidden layers. The number of neurons ($N$) in these layers served as the primary independent variable. Training was conducted using an adaptation of the AlphaGo Zero ($\alpha$GZ) algorithm, which integrates Deep Neural Networks with Monte Carlo Tree Search (MCTS). The networks learned entirely through self-play over 40,000 games, and their performance was evaluated using the Elo rating system.
Results
The primary finding of this research is the discovery of a "Ratio Law" governing competitive success. The win rate of Network A ($N_A$ neurons) against Network B ($N_B$ neurons) follows a remarkably simple mathematical relationship:
$$P_A = \frac{N_A}{N_A + N_B}$$
As illustrated in Figure 1a, the experimental results (black lines) match the theoretical prediction (red lines) to an astonishing degree. This relationship highlights two distinct regimes of investment: when $N_A \ll N_B$, success increases linearly with the addition of neurons; however, as $N_A$ approaches $N_B$, agents face diminishing returns.
Furthermore, the Elo rating of the networks was found to scale logarithmically with the number of neurons, following the relation $y = 400 \log(N) + \text{const}$.
This scaling, shown in Figure 2, provides a bridge between computational complexity and competitive standing, confirming that "intelligence" in adversarial tasks is a relative metric.
Discussion and Conclusion
The results suggest that in competitive environments, an agent does not need to achieve absolute mathematical perfection. Instead, success is determined by relative merit—being slightly better than the opponent. The transition from linear growth to diminishing returns provides a theoretical basis for why biological organisms do not evolve infinitely large brains; the metabolic cost eventually outweighs the marginal competitive gain.
While the "Ratio Law" is robust for the 16-item knapsack problem, limitations remain. Future research should investigate whether these scaling laws hold for larger problem sizes or different architectures, such as Transformers. Furthermore, explicitly modeling metabolic costs would allow for the determination of an "optimal" brain size for a given environment, further bridging the gap between AI and evolutionary biology.
Summary updated