Cooperative Group Optimization (CGO)

• Cooperative Group Optimization is defined by dual-memory agents that integrate private and social knowledge to boost optimization efficiency.
• It employs a portfolio of embedded search heuristics and explicit memory protocols to hybridize and generalize established metaheuristic frameworks.
• Empirical results in sequencing, TSP, and constrained benchmarks demonstrate CGO's rapid convergence, sustained diversity, and robust search capabilities.

Cooperative Group Optimization (CGO) is a general optimization framework inspired by the dynamics of natural cooperative groups, such as animal flocks, insect colonies, or human teams. CGO organizes a population of interacting agents, each possessing private and shared memory, to collectively solve hard optimization problems through mixed modes of individual and social learning. The agents' cooperation is orchestrated via explicit memory protocols, agent portfolios of embedded heuristics, and explicit mechanisms for information exchange and hybridization. CGO encompasses a design space that subsumes and generalizes various population-based heuristics, including ant systems, evolutionary algorithms, and algorithm portfolios, enabling robust search in discrete, combinatorial, and constrained optimization landscapes (Xie et al., 2018, Xie et al., 2018, Xie, 2018).

1. Theoretical Foundations and General Structure

Cooperative Group Optimization is constructed upon the dual-memory paradigm where each agent maintains an individual memory (often denoted $M_A$) for private learning, and has access to a group or social memory ($M_S$) shared across the collective. The interaction protocol is formalized as a multi-agent Markov process, where:

• Each agent $k$ is defined with a private memory $M_A(k)$, generative buffer $M_G(k)$, and update buffer $M_{BA}(k)$.
• The group maintains a social long-term memory $M_S$ and social update buffer $M_{BS}$, typically via an interactive center IC.
• A facilitator controls the search landscape representation $F_R = \langle S, R_M, AUX \rangle$, handles evaluation, and manages global solution records (Xie et al., 2018, Xie et al., 2018).

Agent behaviors cycle through (i) generation of new solutions by mixing $M_A(k)$ and $M_S$0, (ii) submission of candidate solutions to both memories, and (iii) update of both private and social memories using specified update rules (e.g., greedy, tournament, or probabilistic inclusion). This explicit memory protocol enables persistent diversity (through individual memories) while leveraging accelerated convergence via social learning cues.

2. Memory Protocols and Portfolio Search Heuristics

The CGO system operationalizes agents equipped with portfolios of Embedded Search Heuristics (ESHs), modularized operators parameterized by input/output memory access. Each ESH is defined as a parametrized search operator $M_S$1 requiring:

• An ordered list of input chunks $M_S$2, sourced from either $M_S$3 or $M_S$4.
• An output chunk $M_S$5, exported into the agent's generative buffer $M_S$6.

Agents' meta-managers ($M_S$7) orchestrate portfolios by probabilistically selecting among ESHs to operate in each cycle. Hybridization at the low level is achieved by constructing shared memory update lists, allowing in-place portfolio mixing at the chunk level without further algorithmic modification. Cooperative interaction between ESHs is defined when their update lists $M_S$8 overlap, creating hybrid CGO cases where solution components or statistics propagate between otherwise standalone heuristics (Xie et al., 2018).

The framework specifies these components through a multilayered script, assembling problem specification, memory protocol, ESH definitions, and portfolio meta-management, all supporting compositional expansion. The updatable graph derived from the script describes the dependency structure of memory updates and is rooted in generative buffer chunks $M_S$9.

3. CGO Instantiations: RTGO and CGO-AS

CGO has been concretely instantiated in multiple domains:

• Round-Table Group Optimization (RTGO): Developed for discrete sequencing problems such as the flowshop scheduling problem (FSP) and quadratic assignment problem (QAP), RTGO arranges agents in sessions around a logical round table. Each agent maintains a base idea (individual memory) and receives a social idea from its neighbor. Solution search proceeds by recombining base/social ideas using Socially Biased Combination (SBX) recombination schemes followed by problem-specific local search (LS), e.g., insertion-based for FSP or exchange-based for QAP. Key parameters include the number of agents, session count, and recombination/local search selection. Benchmark results on Taillard’s FSP and QAPLIB show RTGO achieves lower relative percentage deviation (RPD) and superior or competitive computation time compared to ILS, PSO–VNS, GRASP, and HAS (Xie, 2018).
• CGO with Ants (CGO-AS): This variant adapts classical Ant System (AS) within the CGO paradigm, equipping each ant with explicit individual memory (personal best tour) and shared pheromone memory. Ants construct new tours using a tunable proportion $k$0 of individual and social cues; a segment of the personal best solution is inherited, with the residual built using ACO’s transition rule based on pheromones and heuristics. Mixing $k$1 empirically yields the best performance on TSP, maintaining population diversity and ensuring rapid convergence, outperforming standard MMAS and recent PSO-ACO hybrids on TSPLIB benchmarks (Xie et al., 2018).

4. Algorithmic Workflow and Hybridization Mechanisms

A typical CGO cycle proceeds as follows (Xie et al., 2018, Xie, 2018, Xie et al., 2018):

1. Initialization: Each agent's $k$2 is seeded (e.g., random solutions, personal best selection), and social memory $k$3 is initialized (e.g., pheromone matrix, population archive).
2. Main Loop (for $k$4 cycles):
   • For each agent:
     • Select an ESH from the portfolio (via $k$5 meta-manager).
     • Generate a candidate solution by combining $k$6 and $k$7, possibly using recombination schemes (e.g., SBX, mixed-tour construction).
     • Optionally apply local search for refinement (problem-specific).
     • Submit candidate to generative buffer, and update $k$8 if the candidate improves over current memory.
   • After all agents: aggregate updates in $k$9 (e.g., pheromone update, public pool update).
   • Facilitator updates problem landscape or search statistics as needed.
3. Termination: After $M_A(k)$0 cycles (or if a convergence/stopping criterion is met), return the best solution observed by any agent.

Hybridization is systematically realized via customized ESH update lists and script-level dependency graphs, enabling seamless integration of distinct heuristics (e.g., DE, PSO, SCO) in a single search process.

5. Performance Evaluation and Empirical Properties

CGO’s empirical efficacy has been validated across several benchmark domains:

• RTGO on FSP (Taillard’s set UB05) with macro-SBX and insertion-based LS ($M_A(k)$1) achieves mean RPD $M_A(k)$2 0.36% in 3.8 s, outperforming ILS (RPD 1.25%, 11 s) and PSO–VNS (RPD 0.41%, 104 s), and on QAP (QAPLIB), RTGO with SBX+exchange-LS ($M_A(k)$3) achieves RPD $M_A(k)$4 0.23%, matching or surpassing GRASP and HAS (Xie, 2018).
• CGO-AS on TSP demonstrates that mixed individual/social cue utilization ($M_A(k)$5) consistently outperforms both pure-social and pure-individual ACO variants in RPD, success rate, and convergence speed, holding robustly across TSP instances from N=51 up to N=1577 (Xie et al., 2018).
• On constrained optimization (CEC’06 G01–G13), CGO with hybrid portfolios (#DESC-I) solves all 13 instances in orders-of-magnitude fewer evaluations than classical DE, ES, and PSO variants (Xie et al., 2018).

CGO’s key strengths are population diversity maintenance (through individual memory), rapid learning acceleration (via collective social memory), and the flexibility to encode hybrid portfolios and landscape adaptation.

6. Parameterization, Tuning, and System Architecture

The effectiveness of CGO depends on principled parameter choices:

• Memory Architecture: Determined by script specification, comprising M_A/M_S chunk types, update rules, and updatable dependency graphs.
• Portfolio Composition: Size, selection weights, and cooperation pattern of ESHs are user-tunable.
• Mixing Parameters: For variants such as CGO-AS, the proportion $M_A(k)$6 controlling individual-to-social cue ratio; for RTGO, SBX operator selection and parallel trial count.
• Population Size: Increasing agent number $M_A(k)$7 or ant count $M_A(k)$8 improves exploration but increases computational load; empirically moderate values ($M_A(k)$9) yield robust results (Xie et al., 2018, Xie et al., 2018).
• Landscape Adaptation: Facilitator-implemented rules (e.g., constraint relaxation) permit on-the-fly adjustment of penalty thresholds, feasibility regions, and other problem encodings.

A multi-layered script structure separates framework, toolbox, and algorithmic assembly, facilitating the inclusion and rapid prototyping of heterogeneous search heuristics without code modification (Xie et al., 2018).

7. Applications, Limitations, and Outlook

CGO has been successfully applied to classic sequencing, assignment, TSP, and constrained optimization benchmarks. Its formalism generalizes several established metaheuristics, providing a unifying framework for hybridization and memory-driven cooperation.

Limitations include the lack of formal finite-time global convergence guarantees (as is typical with population-based metaheuristics), and dependence on appropriate parameter and portfolio configuration for each problem domain (Xie et al., 2018, Xie, 2018). Scaling issues may arise with very large agent populations or when inter-agent communication cost exceeds benefit.

A plausible implication is that further research on automated ESH portfolio design, memory structure selection, and landscape-adaptive facilitators could extend the applicability of CGO to broader and more complex problem classes, including dynamic and high-dimensional optimization scenarios.

---

References:

• (Xie et al., 2018) "Cooperative Group Optimization with Ants (CGO-AS): Leverage Optimization with Mixed Individual and Social Learning"
• (Xie et al., 2018) "A Cooperative Group Optimization System"
• (Xie, 2018) "Round-Table Group Optimization for Sequencing Problems"