Parallel Minority Game

• The Parallel Minority Game is a multi-agent model where each agent selects between two fixed choices and benefits by joining its local minority.
• Its overlapping decision structure couples multiple pairwise minority games, complicating global optimization through induced frustration and nontrivial scaling behavior.
• Analysis of various stochastic switching strategies reveals insights into frozen states, variance reduction, and active–absorbing phase transitions.

The Parallel Minority Game (PMG) is a multi-agent extension of the Minority Game in which there are many available choices, but each agent is restricted to exactly two predetermined alternatives. At each discrete time step, all agents act in parallel, choosing whether to stay at their current option or switch to their other assigned option. The local payoff rule remains the Minority Game rule: an agent benefits by occupying the lesser-populated of its own two choices. The central difficulty is that these pairwise minority decisions are not independent, because different agents’ pairs overlap on common locations; the PMG is therefore a coupled collection of minority games played in parallel, and its global optimization problem is substantially harder than in the two-choice Minority Game (Biswas et al., 2020, Biswas et al., 6 May 2026).

1. Definition and formal structure

The PMG generalizes the standard Minority Game from two global choices to a setting with $D>2$ or $M>2$ total choices, depending on notation, while preserving binary decision-making at the agent level (Biswas et al., 2020, Biswas et al., 6 May 2026). In one common formulation, there are $N$ agents and $D$ sites; each agent $i$ is assigned exactly two distinct, fixed options, denoted for example by $(x_i,y_i)$ or $(\ell_i^{(1)},\ell_i^{(2)})$, and at each time occupies one of them (Tyagi et al., 28 Dec 2025, Tyagi et al., 21 May 2026). All agents update synchronously.

The payoff rule is local and personalized. In the formulation of “Local and global optimization in Parallel Minority Games,” agent $i$ receives a positive payoff if it occupies the lesser-populated of its two possible choices at that time step (Biswas et al., 6 May 2026). The paper writes the local payoff as

$$\text{Payoff}_i(t) = \begin{cases} +1 & \text{if agent } i \text{ is in the minority of its two options at time } t\ 0 & \text{otherwise.} \end{cases}$$

This local minority criterion differs from the global objective of uniform occupancy across all choices.

A standard global observable is the population fluctuation or variance across choices. Using the notation of several PMG papers,

$$\sigma^2(t)=\frac{1}{D}\sum_{k=1}^{D}\bigl(n_k(t)-\langle n\rangle\bigr)^2, \qquad \langle n\rangle=\frac{N}{D},$$

or equivalently, with $M>2$0 choices,

$M>2$1

Here $M>2$2 or $M>2$3 is the occupancy of a choice, and $M>2$4 is the average occupancy per site (Vemula et al., 2 Sep 2025, Biswas et al., 6 May 2026). Global optimization seeks to minimize this variance by distributing agents as uniformly as possible.

Two limiting cases are explicit in the literature. When $M>2$5, the PMG reduces to the standard Minority Game (Biswas et al., 2020). When agents can choose among all resources rather than being restricted to two personal alternatives, the system approaches the Kolkata Paise Restaurant-type setting discussed in the PMG literature (Biswas et al., 2020).

2. Coupled competition and the tension between local and global optimization

A defining feature of the PMG is overlap. Each pair of assigned choices can be viewed as a “mini-MG,” but different mini-games are coupled because agents associated with distinct pairs can still meet at the same location (Vemula et al., 2 Sep 2025). Populations at each location are influenced by all agents present there, not only by agents whose personal pair is that location pair. This overlap creates additional competition and raises the complexity of fluctuation reduction relative to the standard Minority Game (Vemula et al., 2 Sep 2025).

The literature distinguishes sharply between local and global optimization. Global optimization demands uniform population over all choices, whereas local optimization asks that each agent be on the less crowded side of its own personal pair (Biswas et al., 6 May 2026). In the two-choice Minority Game these objectives coincide, but in the PMG they generally compete (Biswas et al., 6 May 2026). This is a central structural distinction rather than a secondary modeling detail.

Several observables have been introduced to track this competition. In addition to $M>2$6, “Efficient strategy for Parallel Minority Games” defines the normalized number of agents currently in the majority of their own two-choice comparison,

$M>2$7

and the fraction of excess crowd,

$M>2$8

where $M>2$9 is the Heaviside function (Vemula et al., 2 Sep 2025). These quantities vanish only under ideal balancing conditions.

This structure also clarifies a common misconception: more complete local information does not automatically improve global coordination. In the PMG, actions that are locally rational can perturb other overlapping pairs and induce new crowding elsewhere (Biswas et al., 6 May 2026). The literature repeatedly emphasizes that the PMG is not simply a larger-choice version of the standard Minority Game, but a distinct coupled optimization problem (Biswas et al., 2020, Biswas et al., 6 May 2026).

3. Microscopic decision rules and stochastic strategies

A substantial strand of PMG research studies stochastic switching rules. In “Efficient strategy for Parallel Minority Games,” three decision rules were analyzed (Vemula et al., 2 Sep 2025). Strategy A is global-average-guided: $N$0 It uses the global mean occupancy and the current occupancy of the agent’s present location, but does not compare against the alternate location.

Strategy B is the instantaneous two-choice minority strategy: $N$1 An agent switches only if its current location is more crowded than its alternate (Vemula et al., 2 Sep 2025). This is presented as analogous to the best stochastic strategy for the standard Minority Game.

Strategy C is memory-based and uses the population seen on the last visit to the alternate location: $N$2 This asynchronous, last-visit rule yields the lowest variance among the strategies studied in that paper, but also produces frozen long-time states (Vemula et al., 2 Sep 2025).

A closely related comparison appears in “Local and global optimization in Parallel Minority Games,” where four information levels were examined (Biswas et al., 6 May 2026). These range from majority/minority information only, to current-location population only, to global average only, and finally to full information about current and alternate populations. The paper reports that the Level 2 strategy—current location population only, with a guessed alternate occupancy—

$N$3

achieves the lowest $N$4 and the highest minority fraction $N$5 among the strategies considered (Biswas et al., 6 May 2026).

The two studies converge on a nontrivial conclusion: partial information can outperform full instantaneous information in terms of global fluctuation reduction (Vemula et al., 2 Sep 2025, Biswas et al., 6 May 2026). In the latter paper, increasing the information level to full knowledge of both populations actually worsens global optimization outcomes because agents overreact and crowd previously less-occupied sites (Biswas et al., 6 May 2026). This does not imply that information is detrimental in general; rather, it indicates that in the PMG the interaction between information, synchrony, and overlap is structurally delicate.

The 2020 PMG paper also introduced a stochastic switching rule adapted from the Minority Game. For an agent $N$6 at location $N$7 with alternate $N$8, when $N$9,

$D$0

while minority agents do not switch (Biswas et al., 2020). This was shown to drive the system rapidly toward near-optimal occupancy patterns, subject to the PMG’s intrinsic constraints.

4. Frozen states, frustration, and scaling behavior

One of the most persistent empirical findings in the PMG literature is the emergence of frozen configurations. Under the memory-based last-visit strategy, agents may stop switching because their information about the alternate site is outdated; they remain on the current site even when they are in the majority of their own two options (Vemula et al., 2 Sep 2025). The system then locks into a static configuration with improved but nonminimal variance.

The same paper reports a finite-size scaling form for Strategy C,

$D$1

together with the saturation-time scaling

$D$2

Thus variance reduction improves with system size, but the time to freezing also grows with system size (Vemula et al., 2 Sep 2025).

A virtual random-walk representation was also used to characterize the dynamics. The history of an agent is mapped to

$D$3

with $D$4 when the agent is at one location and $D$5 at the alternate. Strategies A and B produce diffusive walks, $D$6, whereas Strategy C produces ballistic behavior, $D$7, reflecting freezing at a location (Vemula et al., 2 Sep 2025).

The literature interprets these frozen states in terms of frustration. Overlapping constraints imply that a move relieving crowding at one site can worsen crowding elsewhere, and many local minima appear (Vemula et al., 2 Sep 2025, Tyagi et al., 21 May 2026). “Local and global optimization in Parallel Minority Games” explicitly states that the PMG dynamics are akin to spin glasses and notes that annealing can sometimes help escape frozen, globally suboptimal configurations (Biswas et al., 6 May 2026).

A further subtlety is that frozen states are not identical to globally optimal states. The long-time configurations reached by the most efficient stochastic rules generally retain residual variance (Vemula et al., 2 Sep 2025). This distinguishes the PMG from coordination problems in which stochastic optimization can achieve the theoretical minimum variance.

5. Exact spin-glass mapping

The analogy to spin glasses became an exact equivalence in “Spin Glass Mapping of the Parallel Minority Game” (Tyagi et al., 21 May 2026). In that work, the PMG with $D$8 agents and $D$9 locations is mapped to an Ising spin glass in the mean-field limit.

Each agent’s binary choice is encoded by an Ising variable

$i$0

The occupation number of location $i$1 is written as

$i$2

where $i$3 counts agents for whom $i$4 is an available choice and $i$5 (Tyagi et al., 21 May 2026).

The global cost function is the population variance,

$i$6

Expanding $i$7 in the spin variables yields

$i$8

with

$i$9

The couplings $(x_i,y_i)$0 are quenched random couplings, the $(x_i,y_i)$1 are random fields, and the minimization of PMG variance is exactly equivalent to the search for the ground state of this Hamiltonian (Tyagi et al., 21 May 2026).

This result sharpens earlier qualitative claims. In the 2025 strategy paper, the spin-glass comparison was framed as an analogy based on frustrated minima and freezing (Vemula et al., 2 Sep 2025). In the 2026 mapping paper, the correspondence is exact: PMG optimization becomes a Sherrington–Kirkpatrick-type ground-state problem (Tyagi et al., 21 May 2026). The paper further emphasizes that both positive and negative values of $(x_i,y_i)$2 occur, creating cooperative and antagonistic tendencies simultaneously and thereby producing frustration.

A plausible implication is that PMG instances inherit the computational hardness associated with rugged spin-glass landscapes. The mapping paper explicitly states that, as in SK spin glasses, the problem is generally NP-hard and involves exponentially many configurations (Tyagi et al., 21 May 2026).

6. Non-equilibrium phases and active–absorbing transitions

Beyond optimization, the PMG has also been studied as a non-equilibrium dynamical system with absorbing-state transitions. “Active-Absorbing Phase Transitions in the Parallel Minority Game” treats the PMG as a synchronous adaptive multi-agent model with control parameter

$(x_i,y_i)$3

and critical value $(x_i,y_i)$4 (Tyagi et al., 28 Dec 2025). For $(x_i,y_i)$5, the system can settle into a frozen state without competition; for $(x_i,y_i)$6, persistent overcrowding implies sustained activity.

The paper defines two order parameters: $(x_i,y_i)$7 which measures total crowding above capacity, and

$(x_i,y_i)$8

the fraction of overcrowded sites (Tyagi et al., 28 Dec 2025). Their steady-state limits $(x_i,y_i)$9 and $(\ell_i^{(1)},\ell_i^{(2)})$0 characterize the phase transition.

For instantaneous, memoryless rules such as

$(\ell_i^{(1)},\ell_i^{(2)})$1

the paper reports mean-field directed-percolation scaling with exponents

$(\ell_i^{(1)},\ell_i^{(2)})$2

consistent with mean-field DP (Tyagi et al., 28 Dec 2025). By contrast, threshold-based activation rules produce a distinct non-mean-field universality class, with

$(\ell_i^{(1)},\ell_i^{(2)})$3

and systematic failure of mean-field directed-percolation dynamical scaling (Tyagi et al., 28 Dec 2025).

The interpretation given is that thresholding acts as a relevant perturbation. In the language of the paper, minimal cognitive inertia at the microscopic level can alter the macroscopic universality class (Tyagi et al., 28 Dec 2025). This result extends the PMG beyond static optimization and frozen-state analysis into the theory of absorbing transitions in adaptive systems.

The comparison with frozen states should be kept precise. The absorbing states of the active–absorbing transition are static no-switching states in the dynamical sense (Tyagi et al., 28 Dec 2025). Frozen suboptimal states in optimization studies are also static, but the two literatures emphasize different aspects: one focuses on order parameters and universality, the other on variance minimization and frustration (Vemula et al., 2 Sep 2025, Tyagi et al., 28 Dec 2025).

7. Generalizations, applications, and relation to neighboring models

The PMG has served both as a model in its own right and as a special case within broader game-theoretic constructions. The most explicit generalization in the supplied literature is the Hypergraph Minority Game with Local Hyperedge Payoffs (HMG-L), where overlapping competitions are organized by hyperedges of a static hypergraph (Zhu et al., 2 Jul 2026). That work states that the PMG is recovered when all hyperedges are dyadic, $(\ell_i^{(1)},\ell_i^{(2)})$4, and the interaction structure matches the contest graph induced by the PMG architecture (Zhu et al., 2 Jul 2026). In that dyadic limit, the sparse-regime critical surface

$(\ell_i^{(1)},\ell_i^{(2)})$5

reduces to

$(\ell_i^{(1)},\ell_i^{(2)})$6

which the paper identifies with the PMG or networked-MG result (Zhu et al., 2 Jul 2026).

The PMG has also been applied to movement optimization during an epidemic. In the 2020 application paper, each mobile agent chooses between two assigned regions and attempts to move toward the less infected one (Biswas et al., 2020). The movement probability is written as

$(\ell_i^{(1)},\ell_i^{(2)})$7

when the current region $(\ell_i^{(1)},\ell_i^{(2)})$8 is more infected than the alternate $(\ell_i^{(1)},\ell_i^{(2)})$9. Within the model studied there, coordinated stochastic movement lowers the total infected population relative to random movement and can perform comparably to prolonged movement stoppage, subject to the paper’s assumption that infection risk during travel is not considered (Biswas et al., 2020).

The PMG also sits within a broader family of minority-type models. The standard Minority Game provides the binary global-choice limit (Biswas et al., 2020, Biswas et al., 6 May 2026). Multi-resource minority-game models study many resources without the PMG’s fixed two-option restriction, and exhibit phenomena such as grouping of strategies (Huang et al., 2012). The co-action minority-game literature addresses fully rational coordination in the two-choice case and is relevant mainly as a contrast: in those models, rational decentralized strategies can reach highly efficient cyclic states (Dhar et al., 2011, Rajpal et al., 2018, Sasidevan et al., 2012), whereas in the PMG the overlap structure generally prevents the coincidence of local and global optima (Biswas et al., 6 May 2026).

Taken together, the PMG literature presents a coherent picture. The model is a many-choice, binary-access minority game whose essential novelty lies in overlap-induced coupling. That coupling generates a competition between local minority payoffs and global variance minimization, produces frozen frustrated states under efficient stochastic rules, supports absorbing-state critical behavior with rule-dependent universality, and admits an exact mapping to a mean-field Ising spin glass (Vemula et al., 2 Sep 2025, Tyagi et al., 28 Dec 2025, Tyagi et al., 21 May 2026). This suggests that the PMG occupies an intermediate position between resource-allocation games, non-equilibrium phase-transition models, and disordered optimization problems.