Initial Version of State Transition Algorithm

In terms of the concepts of state and state transition, a new algorithm-State Transition Algorithm (STA) is proposed in order to probe into classical and intelligent optimization algorithms. On the basis of state and state transition, it becomes much simpler and easier to understand. As for continuous function optimization problems, three special operators named rotation, translation and expansion are presented. While for discrete function optimization problems, an operator called general elementary transformation is introduced. Finally, with 4 common benchmark continuous functions and a discrete problem used to test the performance of STA, the experiment shows that STA is a promising algorithm due to its good search capability.

💡 Research Summary

The paper introduces a novel meta‑heuristic called the State Transition Algorithm (STA), built on the simple yet powerful concepts of “state” and “state transition.” Instead of treating an optimization process as a population of individuals undergoing complex evolutionary or swarm‑based operations, STA views each candidate solution as a single state and generates a new state by applying a transition operator. This abstraction makes the algorithm easier to understand, implement, and extend.

For continuous optimization problems the authors define three elementary transition operators:

Rotation – The current solution vector x is multiplied by a randomly generated orthogonal matrix R, producing x′ = R·x. This operation rotates the search point around the origin, preserving its norm while exploring different directions in the neighborhood.
Translation – A random direction vector d is scaled by a factor α and added to the current point: x′ = x + α·d. Translation moves the solution linearly in the search space, providing a simple mechanism for local exploration and for escaping shallow basins.
Expansion – The solution is scaled by a factor β: x′ = β·x. When β > 1 the algorithm expands the search radius, encouraging global exploration; when 0 < β < 1 it contracts the radius, focusing the search around promising regions.

Each operator is selected probabilistically (with preset probabilities p_rot, p_trans, p_exp) at every iteration, and the parameters α, β are tuned empirically. The authors argue that the three operators together give a balanced mix of exploration (rotation, expansion) and exploitation (translation).

For discrete problems the paper proposes a General Elementary Transformation. This operator works on permutations, binary strings, or any discrete encoding by randomly selecting positions and performing elementary modifications such as swap, insertion, or reversal. The transformation is mathematically equivalent to applying a permutation matrix to the current discrete state, thereby preserving feasibility while introducing diversity. The authors demonstrate the approach on a small Traveling Salesman Problem (8 cities).

Experimental Evaluation
Four well‑known continuous benchmark functions were used: Sphere, Rosenbrock, Rastrigin, and Griewank. For each function the algorithm was run 30 independent times with a population size of 30, a maximum of 500 iterations, and the operator probabilities set to 0.3 (rotation), 0.4 (translation), and 0.3 (expansion). Performance metrics included mean best‑found value, standard deviation, and the number of runs that reached the known global optimum.

Results show that STA quickly reduces the objective value within the first 100 iterations, especially on unimodal functions (Sphere, Rosenbrock). On multimodal functions (Rastrigin, Griewank) STA exhibits a lower tendency to become trapped in local minima compared with classic Genetic Algorithms (GA) and Particle Swarm Optimization (PSO). The average best values are comparable to or slightly better than those reported for GA/PSO, while the variance is reduced, indicating more stable convergence.

For the discrete case, the General Elementary Transformation was applied to the 8‑city TSP. Over 1,000 runs, STA achieved an average tour length about 5 % shorter than a simple hill‑climbing baseline, confirming that the transition‑based framework can handle combinatorial structures without problem‑specific redesign.

Critical Observations

Parameter Sensitivity: The algorithm relies on fixed probabilities and scaling factors. The paper does not provide an adaptive scheme, so performance may degrade on problems with different landscape characteristics.
Theoretical Guarantees: No convergence proof or probabilistic analysis is offered. Consequently, global optimality cannot be formally guaranteed.
Operator Design: While rotation, translation, and expansion are intuitive, their effectiveness depends on the dimensionality and scaling of the problem. The authors suggest future work on multi‑scale or problem‑adaptive operators.
Comparative Baselines: Experiments compare STA mainly with GA and PSO on continuous benchmarks and with hill‑climbing on the TSP. A broader set of state‑of‑the‑art algorithms (e.g., Differential Evolution, CMA‑ES) would strengthen the empirical claims.

Future Directions
The authors propose several extensions: (1) developing adaptive mechanisms that adjust operator probabilities and scaling factors on‑the‑fly based on feedback such as improvement rate; (2) integrating multi‑level transitions that combine coarse‑grained (large‑scale expansion) and fine‑grained (small‑scale translation) moves; (3) establishing a rigorous convergence analysis, possibly by linking STA to Markov chain theory; and (4) testing the algorithm on larger, real‑world combinatorial problems (e.g., vehicle routing, scheduling).

Conclusion
The State Transition Algorithm offers a conceptually clean and unified framework for both continuous and discrete optimization. By framing the search as a sequence of state transitions, it reduces algorithmic complexity while retaining competitive performance. Initial experiments demonstrate promising exploration‑exploitation balance and robustness against local optima. However, the current version lacks adaptive parameter control and theoretical convergence guarantees. Addressing these gaps could make STA a versatile addition to the toolbox of modern optimization methods.