Criticality → collective intelligence

Research summary

One paper, summarised for reference. Criticality triggers the emergence of collective intelligence in groups — Ilario De Vincenzo, Ilaria Giannoccaro, Giuseppe Carbone & Paolo Grigolini, Phys. Rev. E (2017) 96, 022309. doi:10.1103/PhysRevE.96.022309.

The claim

A group solving a hard combinatorial problem develops collective intelligence — the ability to settle on high-fitness decisions — precisely at a critical transition, where a large amount of information “leaks” from the problem’s fitness landscape into the group’s shared state. The same Grigolini-lineage criticality idea as the swarm work, but on a decision / optimisation task rather than spatial flocking.

The model (a decision-making model, DMM)

$M$ interacting agents each hold an opinion vector of $N$ binary spins $\sigma_k^j=\pm1$ (agent $k$‘s stance on decision variable $j$), arranged as a multiplex network: within each “decision layer”, spins interact by an Ising energy (social conflict); layers are coupled through each agent’s fitness. The Hamiltonian splits social energy from fitness:

$$H(s) = E(s) - \rho V(s) = -\tfrac{1}{2}J\, A\, s\cdot s - \rho V(s),$$

with $E$ the Ising social-interaction (disagreement) energy, $V$ the configuration’s fitness, and $\rho$ weighting the two. A continuous-time Markov chain evolves opinions; each flip rate is a product of two drives — an Ising–Glauber term (consensus-seeking, minimising conflict) and a Weidlich exponential term (self-interest, favouring fitness-improving flips):

$$w(s_l\to s_l') = \tfrac{1}{2}\!\left[1 - s_l\tanh\!\Big(\tfrac{\beta J}{\langle\kappa\rangle}\textstyle\sum_h A_{lh}s_h\Big)\right]\exp\{\beta'\,\Delta V\}.$$

Three control parameters result: the social interaction strength $\beta J$ (in units of social temperature), the self-confidence $\beta'$ (trust in one’s own expertise), and the knowledge level $p$ (probability an agent knows a given fitness contribution). The landscape is a Kauffman NK model whose complexity $C = K+1+\log_2 N$ is tuned by $K$; each agent perceives only the fraction $p$ of it that it knows. Group decisions are read out by majority rule, with consensus $\chi$ the normalised spin–spin correlation.

The critical front, and the central result

Phase diagrams over $(\beta J,\beta')$ show a critical transition front where the stationary group fitness $V_\infty$ and consensus $\chi_\infty$ jump concurrently from low to high (a disordered, low-consensus phase → an ordered, aligned phase). The paper’s key quantity is the mutual information between fitness and consensus,

$$\mathrm{MI}(\chi_\infty, V_\infty),$$

read as a proxy for how much the group’s shared state has come to “know” the landscape. It is small everywhere except at the critical front, where it is maximal — so at criticality information leaks from the fitness landscape into the group, improving its exploration and triggering the emergence of collective intelligence.

The knobs

• Self-confidence has an optimum $\beta'_{\min}$. Too low ($\beta'=0$ is pure consensus-seeking, blind to fitness) is inefficient; too high makes agents refuse to change their minds even when wrong, suppressing exploration → groupthink (high consensus, low fitness, far from criticality). An intermediate $\beta'_{\min}$ minimises the critical $(\beta J)_C$ and maximises performance.
• Group size $M$ cuts both ways — larger $M$ helps on the low-confidence (thick) branch of the front but hurts on the high-confidence (thin) branch, echoing the empirical “are big teams better?” debate.
• Knowledge $p$: critical dynamics survive cognitive limits ($p<1$) — poorly-informed agents ride consensus to follow the better-informed — but very low $p$ needs agents to be less self-confident and trust peers more.
• Leadership: a very self-confident leader is not generally beneficial and can be detrimental; best performance needs followers confident enough to keep exploring.

References

De Vincenzo, Giannoccaro, Carbone & Grigolini (2017), Phys. Rev. E 96, 022309 · Carbone & Giannoccaro (2015), Eur. Phys. J. B 88, 339 · Glauber (1963), J. Math. Phys. 4, 294 · Kauffman & Levin (1987), J. Theor. Biol. 128, 11 · Weidlich (1991), Phys. Rep. 204, 1 · Vanni, Lukovic & Grigolini (2011), Phys. Rev. Lett. 107, 078103 · Woolley et al. (2010), Science 330, 686 · Mora & Bialek (2011), J. Stat. Phys. 144, 268.