Mitigating Congestion in Complex Transportation Networks via Maximum Entropy

Mitigating Congestion in Complex Transportation Networks via Maximum Entropy


In this paper, we reveal the relationship between entropy rate and the congestion in complex network and solve it analytically for special cases. Finding maximizing entropy rate will lead to an improvement of traffic efficiency, we propose a method to mitigate congestion by allocating limited traffic capacity to the nodes in network rationally. Different from former strategies, our method only requires local and observable information of network, and is low-cost and widely applicable in practice. In the simulation of the phase transition for various network models, our method performs well in mitigating congestion both locally and globally. By comparison, we also uncover the deficiency of former degree-biased approaches. Owing to the rapid development of transportation networks, our method may be helpful for modern society.


Entropy, as a key concept, plays an important role in many fields such as information theory, statistical physics and so on [1]; [2]. In the study of complex network, the entropy has been used to characterize the properties of network topology [3]; [4] and diffusion process on network [5]. Via application of maximum entropy method, we can optimize the network topology for special function and make prediction of network properties [6]; [7].

Traffic flow, modeling the process of material exchange between cells, packet transfer on the Internet, municipal traffic and so on [8]; [9]; [10], is one of the significant research fields in complex networks. While traffic congestion is the main hindrance to transport network’s normal and efficient functioning [11], some models, such as random walk [12], shortest path [13] and so on, has been built to analyze this phenomenon and studies has been focused on optimizing network topology and improving the traffic capacity to achieve free-flow for these different models [14]; [15]. Most of previous studies concentrate on optimizing the topology of network and designing better routing strategies [16] to improve the efficiency of transportation. These methods, usually needing high cost to change infrastructure or very concrete and global information of the traffic process, are designed for some special cases and maybe uneconomical and unattainable in practice.

Our study is complementary to the above efforts. Here, based on maximum entropy method, we develop an approach to mitigating congestion by allocating limited capacity to nodes reasonably in network with observable information in traffic process. We reveal that we can improve the whole system’s traffic efficiency by just improving a little node capacity without changing network topology. Our method is tested in various models for different networks, performing well in raising traffic efficiency.

Before presenting our approach, we give a brief description of network traffic. At each time step, each node has a probability to generate a packet which has a destination and chooses a path towards its destination in network immediately. And a node can deliver at most packets to the next node of its path at a time step. As the capacity is limited, the packets which arrive to the node earlier have priority to be delivered. A packet is removed from the traffic after reaching its destination.

We suppose that we can get each node’s local information including the frequency of different neighbors it delivers packets to and the number of packets it delivers in traffic. The frequency and number, as local quantities decided both by the traffic dynamical process and by the network topology, is typically easy to be observed and counted in reality. Let us analyze the routing of packets which is certain under specific strategy from a stochastic standpoint. Mathematically speaking, we regard the traffic process on a network composed from nodes as a time invariant ergodic Markov chain with a transition matrix , where the entry is the average frequency of packets to go from node to at a time step satisfying . We introduce entropy rate to characterize this process. An ergodic Markov chain with transition matrix has a unique stationary distribution satisfying


where is the transposed matrix of . And can be computed directly with normalization condition . The dynamical properties of the network can be evaluated by entropy rate that is given by


The entropy rate is a quantity studied in the diffusion process and measure the time density of the average information in a stochastic process. It has been showed that a high entropy rate indicates a large randomness, or easiness of propagating from one node to another [5]; [7]; [17]. Therefore it can also be related to an efficient and free traffic over the network. In that paper, our main goal is to mitigating traffic congestion by maximizing the associated entropy rate function.

A very simple method to mitigate congestion is to improve the capacity of each node as high as possible. However, the capacity is usually limited by various conditions. And an fundamental and significant problem is about how to allocate the limited capacity to a given network to make the system efficient. Next we try to solve this problem by maximizing the entropy rate. The node in network corresponding to the row vector of network’s transition matrix , where , according to the above generation method of matrix , can represent the possibility of a packet at node going to node at a time step in large scale approximately. When congestion occurs in node , we have , that is, the packets have a possibility to wait in node for being delivered at a latter time step. And the direct consequence of improving the node capacity is the decrease of . We denote the as the detention probability of node in this paper, which plays an important role in our model. Suppose the change of is () after being allocated new capacity (i.e. the probability of a packet stay in the node is ). Since the nodes deliver the packets by time-priority which is not associated with the packets’ routing, it is natural to get that the th () component of vector is . Therefore we can view as a vector function of variables . While is composed from , we can get the entropy rate as a function of variable from equations 1 and 2.

It is obvious that a higher node capacity will lead to a lower detention probability . And we need a quantitative relationship to identify the feasible region of . As mentioned above, we suppose we can get the the average number of packets that the node received at each time step is , which can be counted relatively easily in practice or be computed analytically for some special cases [11]. Suppose the change of is . Approximately, we have that


that is, the decrease of detention probability approximately equal to the ratio of the increment of the number of packets a node can deliver per time step and the the average number of packets that the node received at each time step. Then for a given model and limited node capacity budget . we model our problem as follow: for


maximize the function


From equations 4, the feasible region of optimization variable is a convex set. And the entropy function have some nice properties. Many sophisticated methods, such as Monte Carlo algorithms [18], have been developed from such problem before [19], and can be applied directly. Suppose that get its maximum at on the feasible region. We allocate capacity to node . Generally, is not integer and we need to adjust our distribution as close as possible to it. For the case the function is analytic at ,


By the method of Lagrange multipliers, we can range nodes in network by the priority for the case of limited traffic capacity budget approximately, and allocate the budget to the nodes with higher priority preferentially within the feasible region. Incidently, with less information, since


for any initial distribution , can be approximated by , where is the average number of packets in the whole network at each time step. Therefore we can also range the nodes by . As the complexity of solving a system of linear equations with variables is , the distribution strategy can be get at most steps by numerical computation. Further more, with a scale-free (power-law) node-degree distribution [20], most of the real networks, especially communication networks, are sparse. And there are more efficient methods for these sparse networks [21].

Next we solve entropy rate as a function of detention probability for two simple and basic systems analytically explicitly: (i) a star-like system and (ii) a chain-like system [22]. For simplicity, we suppose the detention probability of node is and . (i) For a starlike system composed from nodes as figure 1(a) shows, we have the transition matrix , where and otherwise . Solve the eigenvector of and we have the stationary distribution: where , and the normalization factor . Calculate entropy rate from equations 1 and 2:


(ii) For a chain-like system composed from nodes as figure 1(b) shows, we suppose the transition matrix , where and , otherwise . Similar to the above, we get the stationary distribution , where for , for , and the normalization factor . Then the entropy rate is


For these two systems, we can analyze their entropy rate functions in the neighborhood of dentation probability and apply our method.

We test our method for the congestion model based on reference [11], where the node capacity at the beginning, is a control parameter and is the degree of node . The total number of nodes in the network is . As , the possibility for a node to generate a package at each time step, is increased from zero, two phases can be observed: free flow for small and a congested phase for large . And there is a phase transition at . To simulate the practical congestion process, such as the traffic congestion during peak hours and the congestion of communication during major holiday, we create packets at each time step with some probability distribution for a time period, and stop creating after congestion occurring for a while, then the congestion cease after a time period. As what we have mentioned before, we count the frequency in the process of congestion and apply our method to allocate limited capacity budget to the nodes in network. It is worth mentioned that, different from previous work [11], our method and calculation, based on observation, do not require a known probability distribution of packet generation and the packets’ concrete routing strategy. For comparison, we allocate the same capacity budget randomly and by a degree-biased strategy, where the degree-biased strategy, similar to reference [11], is to increase the control parameter to allocate capacity budget to nodes. Then we create packets in the same way again and compare the traffic efficiency between these methods. Here we use quantity to characterize the traffic efficiency. Here is the velocity of a node being navigating in network.


where is the length (i.e. the number of nodes a packets passed from start position to destination) and is sum of time steps it takes. The indicates the average over all packets. We present our simulation for (i) Cayley tree, (ii) regular network, (iii) random network and (iv) scale-free network in Fig. 1. As it shows, we can see that, for different node capacity budgets, corresponding being lower than others in general, our method generally performs well in improving the traffic efficiency with limited capacity budget.

Figure 1: (Colour On-line) graph, the average speed is shown as a function of for (a) Cayley tree, , , , thus , (b) regular network, , , , (c) random network, , , (d) scale-free network, , , with different capacity allocation methods, where the argument is the ratio between initial capacity sum and improved capacity sum of the whole network, and triangle, plus and circle curves correspond to the simulations of our method, degree-biased method and random method. The initial capacity distribution follows at the beginning. We simulate the process of congestion by keeping for time steps and choosing the generating packages’ destinations randomly.

Former approaches of controlling traffic congestion, especially for heterogenous networks such as scale-free networks [14]; [23], concentrated on optimizing the nodes with high degree. However, as Fig. 2 shows, we find that the importance of nodes in the traffic process is not totaly dependent on the their degree, that is, the nodes with high degree may not more important and need to be allocated more capacity than those with low degree. The nodes with higher priority are fairly independent of their degree. For a fixed degree range, there is a wide spread of priority. Therefore our method might be a modification of the former degree-biased study.

Figure 2: (Colour On-line) Our method identifies the important nodes for limited capacity budget more reliably than the degree-biased method. The nodes are permutated with their priority computed from the equation 6 in the order from highest to lowest on the horizontal axis and the ordinate axis corresponds to their degree . The distribution of is shown via heat graph for (a) random network and (b) scale-free network. The conditions where we get the results from are the same as the conditions in Fig. 1 for random network and scale-free network.

Further more, we reveal that our method can mitigating the congestion globally with local information. In more detail, the transition frequency of each node what we get is the phase of network in a special phase with a . However, as the stationary distribution and entropy rate are global for the dynamical process on complex network, our method may also be effectual on a much larger scale. By simulation, we find that we can optimize a wide range of phases with only the information got in one phase. We allocate capacity budget by our method for a proper phase with a large . And then test the network with different . Similar to the above, we repeat the same process for random strategy and degree-biased strategy as a comparison. It has been showed that [8], after congestion occurring, the number of packets in the network will grow linearly with time. And there is a parameter:


to qualify the transition, where and indicates the average over time windows of . We use this parameter to evaluate the global effect of our optimization for two heterogenous networks: Cayley tree and scale-free network. As Fig. 3 shows, the corresponding of method is lower and the threshold of phase transition is higher than others. Our method has a global optimal effect on mitigating congestion.

Figure 3: (Colour On-line) graph, the parameter is shown as a function of for (a) Cayley tree, , , , , , , and (b) scale-free network, , , , with different optimization methods, where triangle, plus, star curves correspond to the simulations of our method, degree-biased method and original situation.

Introducing maximum entropy rate method into traffic network may also help us assess a traffic network globally and evaluate the gap between it and the optimal state by comparing it to the maximal-entropy random walk [7]; [24], which can be relatively localized easily. Suppose that the practical transition matrix is , and the transition matrix corresponding to maximal-entropy random walk is . Regarding the transition matrix as random variables, we can use Kullback-Leibler divergence to identify the difference between them:


and if and only if . We can conclude that a traffic process with lower Kullback–Leibler divergence is more efficient generally.

Recently, it has been showed that many real networks are temporal networks which has many advantages[25]. As our method, based on the dynamical state of network, is less dependent on the fixed network topology. It maybe suitable and helpful for the optimization of the communication on temporal network since the priority can be computed for every time step. Further more, as the capacity usually changes over time in temporal networks too, our method may enlighten researchers to study how to allocate traffic capacity dynamically on temporal networks to improving its traffic efficiency.

Summarizing, in this paper, we analyze the quantitative relationship between traffic capacity in congestion and entropy rate in complex network and solve it explicitly for star-like system and chain-like system. Based on the principle that a higher entropy rate leads to the easiness of propagating in complex network, we develop a method to mitigating congestion by allocating the limited traffic capacity to nodes rationally. We analyze the complexity of our method, get that it can be completed in at most steps, therefore is feasible. What is more, as it only requires the local and observable information and does not change the topology of network, this method is shown to be low-cost and generally practical. We test our method and compare it to other methods for (i) Cayley tree, (ii) regular network, (iii) random network and (iv) scale-free network amd get that our method is more functional in general. The congestion is considerably mitigated after applying our strategy. Free flows on networks are critical in our world and how to mitigate congestion with limited budget is a fundamental and significant problem. Our proposed a heuristic method for this problem and may be useful for modern society.


  1. Claude Elwood Shannon. A mathematical theory of communication. ACM SIGMOBILE Mobile Computing and Communications Review, 5(1):3–55, 2001.
  2. Linda E Reichl and Ilya Prigogine. A modern course in statistical physics, volume 71. University of Texas press Austin, 1980.
  3. Martin Rosvall, Ala Trusina, Petter Minnhagen, and Kim Sneppen. Networks and cities: An information perspective. Physical Review Letters, 94(2):028701, 2005.
  4. Ginestra Bianconi. Entropy of network ensembles. Physical Review E, 79(3):036114, 2009.
  5. Jesús Gómez-Gardeñes and Vito Latora. Entropy rate of diffusion processes on complex networks. Physical Review E, 78(6):065102, 2008.
  6. Juyong Park and Mark EJ Newman. Statistical mechanics of networks. Physical Review E, 70(6):066117, 2004.
  7. Zdzisław Burda, J Duda, JM Luck, and B Waclaw. Localization of the maximal entropy random walk. Physical review letters, 102(16):160602, 2009.
  8. Alex Arenas, Albert Díaz-Guilera, and Roger Guimera. Communication in networks with hierarchical branching. Physical Review Letters, 86(14):3196, 2001.
  9. Daqing Li, Bowen Fu, Yunpeng Wang, Guangquan Lu, Yehiel Berezin, H Eugene Stanley, and Shlomo Havlin. Percolation transition in dynamical traffic network with evolving critical bottlenecks. Proceedings of the National Academy of Sciences, 112(3):669–672, 2015.
  10. Albert Solé-Ribalta, Sergio Gómez, and Alex Arenas. Congestion induced by the structure of multiplex networks. Physical review letters, 116(10):108701, 2016.
  11. Liang Zhao, Ying-Cheng Lai, Kwangho Park, and Nong Ye. Onset of traffic congestion in complex networks. Physical Review E, 71(2):026125, 2005.
  12. Zoltán Eisler and János Kertész. Random walks on complex networks with inhomogeneous impact. Physical Review E, 71(5):057104, 2005.
  13. K-I Goh, B Kahng, and D Kim. Universal behavior of load distribution in scale-free networks. Physical Review Letters, 87(27):278701, 2001.
  14. Zhe Liu, Mao-Bin Hu, Rui Jiang, Wen-Xu Wang, and Qing-Song Wu. Method to enhance traffic capacity for scale-free networks. Physical Review E, 76(3):037101, 2007.
  15. Roger Guimerà, Albert Díaz-Guilera, Fernando Vega-Redondo, Antonio Cabrales, and Alex Arenas. Optimal network topologies for local search with congestion. Physical Review Letters, 89(24):248701, 2002.
  16. Jon M Kleinberg. Navigation in a small world. Nature, 406(6798):845–845, 2000.
  17. Adam L Berger, Vincent J Della Pietra, and Stephen A Della Pietra. A maximum entropy approach to natural language processing. Computational linguistics, 22(1):39–71, 1996.
  18. Steffen Bohn and Marcelo O Magnasco. Structure, scaling, and phase transition in the optimal transport network. Physical review letters, 98(8):088702, 2007.
  19. Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004.
  20. Albert-László Barabási and Réka Albert. Emergence of scaling in random networks. Science, 286(5439):509–512, 1999.
  21. William F Tinney and John W Walker. Direct solutions of sparse network equations by optimally ordered triangular factorization. Proceedings of the IEEE, 55(11):1801–1809, 1967.
  22. Jianxi Gao, Sergey V Buldyrev, Shlomo Havlin, and H Eugene Stanley. Robustness of a network of networks. Physical Review Letters, 107(19):195701, 2011.
  23. Zonghua Liu, Weichuan Ma, Huan Zhang, Yin Sun, and Pak Ming Hui. An efficient approach of controlling traffic congestion in scale-free networks. Physica A: Statistical Mechanics and its Applications, 370(2):843–853, 2006.
  24. Roberta Sinatra, Jesús Gómez-Gardenes, Renaud Lambiotte, Vincenzo Nicosia, and Vito Latora. Maximal-entropy random walks in complex networks with limited information. Physical Review E, 83(3):030103, 2011.
  25. Aming Li, Sean P Cornelius, Yang-Yu Liu, Long Wang, and Albert-László Barabási. The fundamental advantages of temporal networks. arXiv preprint arXiv:1607.06168, 2016.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description