Collective Annealing by Switching Temperatures: a Boltzmann-type description

The design of effective cooling strategies is a crucial component in simulated annealing algorithms based on the Metropolis method. Traditionally, this is achieved through inverse logarithmic decays of the temperature to ensure convergence to global minima. In this work, we propose Collective Annealing by Switching Temperatures (CAST), a novel collective simulated annealing dynamic in which agents interact to learn an adaptive cooling schedule. Inspired by the particle-swapping mechanism of parallel tempering, we introduce a Boltzmann-type framework in which particles exchange temperatures through stochastic binary interactions. This process leads to a gradual decrease of the average temperature in the system. Numerical results demonstrate that the proposed approach consistently outperforms classical simulated annealing with both logarithmic and geometric cooling schedules, particularly in terms of convergence speed.

💡 Research Summary

This paper introduces a novel metaheuristic optimization algorithm named Collective Annealing by Switching Temperatures (CAST), designed to overcome the limitations of classical Simulated Annealing (SA). Traditional SA relies on a predetermined cooling schedule (e.g., logarithmic or geometric decay) applied uniformly to all sampling agents, which can lead to slow convergence or entrapment in local minima if the schedule is poorly chosen. CAST proposes a decentralized, collective approach where a population of agents, each with its own temperature, interact to learn an adaptive cooling schedule dynamically.

The core innovation of CAST lies in its two-phase iterative process. In the exploration phase, each agent performs a Metropolis-Hastings type random walk in the solution space, where the acceptance probability of worse solutions is governed by its individual temperature. In the interaction phase, pairs of agents stochastically exchange parts of their temperatures based on a binary interaction rule inspired by Parallel Tempering. This rule is designed to cool down agents currently at good positions (low objective function values) and warm up agents at poor positions, thereby promoting exploitation and exploration respectively. Crucially, the interaction parameters are set such that the expected mean temperature of the entire system decreases over time without any external schedule.

To analyze the macroscopic behavior emerging from these microscopic interactions, the authors employ tools from kinetic theory. They derive a Boltzmann-type partial differential equation for the evolution of the probability density function f(x, T, t) of agents having position x and temperature T at time t. This equation combines a linear operator modeling the spatial search and a nonlinear collision operator modeling the temperature exchanges. This kinetic formulation allows for a rigorous mean-field analysis, proving properties like the decay of the average temperature under certain parameter conditions (specifically, when the warming coefficient μ is less than the cooling coefficient λ).

The paper provides extensive numerical experiments on standard benchmark functions. Key implementation choices are discussed, such as the use of Cauchy-distributed noise for proposing new positions, which enables larger jumps and helps escape deep local minima. The results consistently demonstrate that CAST outperforms classical SA with both logarithmic and geometric cooling schedules, particularly in terms of convergence speed and solution quality. The algorithm effectively self-regulates the cooling process, allocating computational resources efficiently by focusing lower temperatures on more promising regions of the search space. The work concludes by positioning CAST as a promising new paradigm that combines the strengths of population-based methods with a sound theoretical foundation derived from kinetic models.