In the year of 2019, the fifth-generation cellular networks (5G) has come into the commercialization phase. With the support of enhanced mobile broadband (eMBB) service, massive-connected machine-type communication (mMTC) service and ultra-reliable low-latency (URLLC) service, 5G is expected to transform cellular network from pure life-based Internet to industry-oriented Internet. Meanwhile, researchers from both academia and industry has been actively looking forward to techniques for sixth-generation cellular networks (6G)111Notably, some in the literature readily used the term 6G while others use terminologies like beyond 5G (B5G) or 5G+. In this article, we assume these terminologies are inter-changeable.. Despite somewhat unclear research directions (e.g., tera-Hertz communications, intelligence-enhanced communications [1, 2], space-air-ground integrated network , etc), 6G has been destined to further bring immerse Internet experience and contribute to an unprecedented convergence between the physical world and the information world as illustrated in Fig. 1.
On the other hand, as a non-nascent concept, collective intelligence (CI) is a form of “universally distributed intelligence, constantly enhanced, coordinated in real time, and resulting in the effective mobilization of skills” . Therefore, being significantly different from swarm intelligence, which is inspired from the collective behavior in biology and usually refers to a general set of centralized algorithms (e.g., ant colony optimization), the basis and goal of CI is “mutual recognition and enrichment of individuals rather than the cult of fetishized or hypostatized communities” . CI has been successfully applied in several well-known examples including Wikipedia and reCaptcha, by enabling humans to interact and to share and collaborate with both ease and speed . For Wikipedia, the Internet gives participating humans the opportunity and privilege to store and to retrieve knowledge through the collective access to these databases . With the development of the Internet and its widespread use, the opportunity to contribute to knowledge-building communities, is greater than ever before. Meanwhile, following the evolving support of rate, latency and number of connections in cellular networks, increasing numbers of things will be surely connected in a more efficient and smarter manner. Therefore, we boldly argue that fostered by 6G, CI will enter into a more advanced level to drive the digitization of information & communication and power the connections among beings and things.
However, several questions naturally arise like that: what role does 6G plays for CI? How can individuals collaborate within the framework of CI? How about the performance after such CI and how can we evaluate the intelligence level of CI? The article tries to discuss these points.
Ii CI and 6G
Along with the rapid urbanization in developing counties, cities have become the preferential habitat and more than 54% of the world’s population is believed to live in cities. In Hangzhou, an eastern provincial capital of China with a population of over 9 million, more than 17 million subscribers were connected to the Internet via cellular networks by smartphones, tablets or laptops in 2016. Meanwhile, a constantly growing number of lower powerful devices like metering sensors and wearable devices urgently join the Internet and facilitate humans’ lives. Furthermore, the weird phenomena where the Internet of vehicles only implies a collection of sensor platforms that provide information to drivers and upload filtered sensor data (e.g., GPS location, road conditions, etc.) to the cloud, is experiencing remarkable changes. As the Level-4 autonomous car soon becomes a reality222https://www.engadget.com/2018/11/02/volvo-baidu-autonomous-cars-china/, a fleet of vehicles in the future needs to exchange their sensor inputs among each other so as to realize prompt delivery of the passengers to destination with maximum safety and comfort and minimum impact on the environment. The Internet of vehicles turns to rely on a combination of eMBB, mMTC and URLLC services rather than the traditionally simple sensor-data acquistion .
However, the widely-existing connection only lays the necessary but insufficient condition for the CI. Usually, CI distinguishes from the pure collaboration by ingredients from four aspects .
Openness: Collective intelligent participants could transcend the hurdle of intellectual property and gain significant improvement from sharing information.
Peering: Collective intelligent participants are “horizontally” equal in terms of both intelligence level and decision-making abilities. The resulting self-organization can be much more effective and efficient at producing the desired results.
Sharing: Sharing implies to allow their fluid exchange of information and collective intelligent participants including both beings and things have the capability to process the exchanged information. The sharing could rely on direct communications or indirect communications (e.g., observing the digital pheromone left on the environment) as well as a more general imitation process.
Acting Globally: Collective intelligent participants get rid of possible barriers (e.g., department) and possess the capability to reach out across their entire network of collaborators to exchange information.
Based on the aforementioned discussions, CI heavily relies on the advancement in communication technology (CT) and artificial intelligence (AI). However, on one hand, the goal to design vertical industry-oriented Internet has not been established until 5G. In particular, 5G begins to classify services into three types (e.g., eMBB, URLLC, mMTC) and separately provision each service using one infrastructure, based on the softwarization and virtualization of network functionalities as well as network slicing techniques to appropriately orchestrate resources. Besides, 5G has introduced the network data analytics function (NWDAF) to better exploit the availability of massive actionable data. In the physical layer, 5G accommodates stringent requirements in terms of data rate and latency, by introducing millimeter-wave (mmW) communications, exploiting massive multiple-input/multiple-output (MIMO) links, and deploying (ultra) dense radio access points . On the other hand, AI has also manifested the astonishing capability in the well-known success of AlphaGo . There is few doubt that everything around us will be very intelligent with the enormous amount of data and information, and pervasive AI will be an integral part of 6G as well [1, 6, 10, 11]. Therefore, the transmission rate is speculated to be significantly improved in 6G with the further reduced latency. Moreover, the static service classification mechanism in 5G potentially becomes more flexible and customizable in 6G.
In summary, Fig. 2 highlights that the advancement in 6G and AI. Moreover, a typical scenario of CI in the future could be the vehicle to everything (V2X). In particular, V2X connects vehicles with other vehicles (i.e., V2V), people with wearable devices or smartphones (i.e., V2P), infrastructure (i.e., V2I) as well as the network (i.e., V2N). AI will plays a crucial role in this collective V2X system, so as to avoid inter-vehicle collision, protect people nearby and realize smart traffic control. Consequently, V2X should be able to sufficiently support the required four ingredients in CI.
Iii The Communicating Language among Collective Intelligent Agents
In order to fulfill the goal of openness and sharing in CI as well as to effectively augment human intelligence, it is incentive to design a communicating language among the agents. As human intelligence is intimately tied to the capacity to understand linguistic expressions and manipulate symbolic expressions , a scientific model of language with computable semantics is highly required. We strongly argue that the information economy metalanguage (IEML)  is an appropriate candidate to achieve such a goal.
As a regular language, IEML adopts an alphabet with six elementary symbols . The left part of Fig. 3 highlights the six elementary symbols and explains its related meaning in natural language. Basically, IEML builds semantic sequences by concatenating symbols . If the length of one sequence is denoted as the number of elementary symbols therein, the sequences of IEML can have only 7 predefined lengths, namely , , , , , , from Layer to Layer . In other words, the length of a sequence is always a power of 3. Generally, a sequence of Layer () is constructed by attaching three sequences of Layer in an end-to-end manner and the three sequences of Layer play different syntactic roles (i.e., substance, attribute and mode in turn). Finally, terminating punctuation marks (i.e., “:”, “.”, “-”, “’”, “,”, “_”, “;”) are used to express the end of sequences from Layer to Layer .
IEML takes advantage of a script to reciprocally transcode the algebra into sequences. The various algebraic expressions that are equivalent are translated by a unique code in the script, and vice versa. During the translation, two operations (i.e., multiplication and addition) are defined. Different from the common meaning in mathematics, the multiplication operation denotes a reversible means to construct sequences for a particular layer from the sequences from its lower layer. Meanwhile, the addition operation means to manipulate sets of sequences by union, intersection or symmetrical difference. As implied in the upper-right part of Fig. 3, the aforementioned transcoding rules allow the script to be computable.
IEML produces and recognizes words, phrases, texts and hypertexts by the IEML dictionary, which is composed of a set of inter-operable keys. The lower-right part of Fig. 3 provides the interpretation of the key *O:M:.**. As the substance of *O:M:.** is *O:.** (i.e., a verb), *O:M:.** could be interpreted as the domain of the action. For example, when the substance is *U:.** corresponding to a virtuality or modality to the action, *U:B:E:.** express the feeling of “want” as *B:.** implies the intentional core of the person . Similarly, it can be observed from Fig. 3, all the dimensions of the action can be found in *O:M:.**.
Notably, IEML is not a programming language. Instead, it could connect with other programming languages to generate and handle texts. The meaning of texts can be analyzed and represented by uniform semantic locators (USLs, i.e., sets of IEML sequences) and their graphs of relationships, which are represented in semantic variables by transforming between concepts and texts. Moreover, all operations based on transformations of semantic variables in IEML are computable by following the properties of finite state machines . Therefore, IEML paves the way to perform translinguistic hypertextual communication after the careful design of an open and universal encyclopedic library, so as to enables creative conversations among collective intelligent agents.
Iv The Achievement by Collective Intelligent Agents
As mentioned above, CI allows agents to cooperate properly to fulfill some complicated task. In this part, we talk about some simulation results to demonstrate the achievement of CI. Concretely speaking, we assume there exist an area with some movable agents and their goal is to form a specific pattern of “” as depicted in Fig. 4. Moreover, the total simulation area can be classified into the labeled area and the unlabeled area, and the labeled area corresponds to the specified shape which agents need to form. For example, in Fig. 4, the total area is occupied by 2828 blocks while the labeled area colored in white needs 119 agents to fill it up. Each movement of the agents will leave some digital pheromone as the medium of an indirect communication among agents while each block in the total area can keep the digital pheromone for a period of time in order to provide guidance for the cooperation of agents. As for the movement of each agent, we set the following rules.
Each agent is supposed to move a block towards one of four directions at a time, that is, UP, DOWN, LEFT, and RIGHT.
Each block in the area can only be occupied simultaneously by an agent.
Each agent can leave the digital pheromone in the current position and sense the amount of digital pheromone within a certain range.
Each agent can identify whether the current position is labeled or unlabeled.
Each agent can sense the existence of neighbors in all directions (i.e., UP, DOWN, LEFT, and RIGHT) and can even communicate with them briefly.
We have provided a CI framework and classified the cooperation between agents into the following five steps:
Selecting active agents: A certain proportion of agents are selected out for moving opportunities. The possibility of each agent being selected is related to the priority of its action itself. Concretely, the agent only with the highest action priority can be regarded as “active” and obtain the moving opportunity. The selection of active agents is also consistent with our intuition that common heterogeneous agents could respond to one task in terms of speed and delay distinctly.
Selecting attractor: Each block with a certain amount of digital pheromone can be regarded as an attractor. In this part, each active agent selects its own attractor in a stochastic manner. The probability of each attractor being selected corresponds to the amount of digital pheromone as well as the distance with the agent.
Moving: After active agents and attractors being determined, each active agent will move one block towards its own attractor. But this agent will stop or select another direction if it encounters the obstruction of neighbors. For example, an agent will move a block towards the “UP” direction if the selected attractor is above this agent. However, if the block to be visited has been already occupied by another agent, this agent will stop or move towards the “LEFT” or “RIGHT” direction.
Leaving digital pheromone: When an agent arrives at a new position, it will leave the digital pheromone in the current position according to the attribute (e.g., labeled or unlabeled) of the area. Specifically, the amount of digital pheromone in this position will be increased if this position belongs to the labeled area while it will be decreased otherwise.
Updating priority: The action priority for a particular agent will be updated according to the information of position as well as the number of surrounding neighbors. During the simulation, a predefined reward table is provided for the updating of action priority.
Fig. 4 provides the simulation results for this typical CI scenario. It can be observed from Fig. 4, along with iterations, the gap between the targeted original image and the shape formed by agents gradually diminishes. After 150 iterations, the similarity between the targeted original image and the shape formed by agents exceeds and demonstrates the effectiveness of CI.
V The Evaluation of Collective Intelligence
Agents could collaborate in different organization and communications manner, resulting into distinct intelligence level of CI. It is crucial yet challenging to quantify the CI, even though there exist many solution for the individual intelligence. Recent efforts towards the evaluation of CI include to build evaluation model , compute system level fault-tolerance  and extract essential features . In this article, considering the flexibility and universality, we focus on the anytime universal intelligence test (AUIT) in , which is primarily concerned about the homogeneous agents with the same type of agents but could be easily extended to the case where heterogeneous agents interact in different ways.
AUIT works in a toroidal grid space with periodic boundaries. In other words, an agent, which moves off one border, will appear on the opposite one. In AUIT, there are objects from a finite set which contains a set of collective intelligent agents () and two moving special objects, Good () and Evil (). The two special objects move in the environment with measurable complexity movement patterns while the movement of agents in is chosen from a finite set of moving actions . The intelligence level of this CI is determined by a predefined reward function mapped from the distance between the evaluated agents to objects and .
As the synergy among the agents is of vital importance to the CI level , AUIT could be easily extended to different communication scenarios like direct communication (i.e., talking), indirect communication (i.e., stigmergy by digital pheromone), and imitation. In particular, different CI agents in AUIT retrieve information distinctly. For example, agents with direct communication will be informed of all agents’ observations exactly, while agents with indirect communication will get others’ observations with a specific bias. Besides, agents with imitation only have the capability to get observations from agents within their observation range.
AUIT could easily perform the sensitivity analysis of the CI level. For different task complexities, AUIT could change the two special objects’ movement patterns and quantify the complexity in terms of Kolmogorov complexity. Meanwhile, AUIT can differ the search space complexity or environmental complexity by tuning the size of the environment (yielding different Shannon entropies) and remain other settings fixed.
In this paper, we have highlighted the CI among connected beings and things, which harnesses the advancement in CT and AI and is assumed to emerge in the 6G era. Following that, we have proposed the IEML as the communicating language among multiple agents to augment human intelligence and also explained the preliminary semantics for IEML. Afterwards, we have demonstrated the achievement by collective intelligent agents through simple indirect communications and have manifested the effectiveness of CI. Finally, we have discussed the feasibility and flexibility of an AUIT model to evaluate the intelligence level of CI.
In fact, there still exists many concerns to be addressed in the future. For example, along with the gradual accumulation of connected intelligent agents, how does the intelligence level of CI vary? What are the reliable realization of emulation platform and involved procedures for CI? We will try to answer these questions in the future.
-  R. Li, Z. Zhao, X. Zhou, G. Ding, Y. Chen, Z. Wang, and H. Zhang, “Intelligent 5g: When cellular networks meet artificial intelligence,” IEEE Wireless Commun., vol. 5, no. 24, pp. 175 – 183, Oct. 2017.
-  E. C. Strinati, S. Barbarossa, J. L. Gonzalez-Jimenez, D. Kténas, N. Cassiau, and C. Dehos, “6g: The next frontier,” arXiv:1901.03239 [cs], Jan. 2019, arXiv: 1901.03239. [Online]. Available: http://arxiv.org/abs/1901.03239
-  Z. Niu, S. Shen, Q.-y. Zhang, and Y.-l. Tang, “Space-air-ground integrated vehicular network for immersive driving experience,” Chinese J. IoT,, vol. 2, no. 1, pp. 17–27, Sep. 2017.
-  P. Levy, The semantic sphere 1: Computation, cognition and information economy. Wiley-ISTE, 2011. [Online]. Available: https://www.amazon.com/Semantic-Sphere-Computation-Cognition-Information/dp/1848212518
-  T. Flew, New media: An introduction. Melbourne: Oxford University Press, 2008. [Online]. Available: http://www.oup.com.au/titles/higher_ed/media_studies/9780195551495
-  M. Gerla, E. Lee, G. Pau, and U. Lee, “Internet of vehicles: From intelligent grid to autonomous cars and vehicular clouds,” in Proc. IEEE WF-IoT 2014, Seoul, South Korea, Mar. 2014, pp. 241–246.
-  D. Tapscott and A. D. Williams, Wikinomics: How mass collaboration changes everything, expanded ed. New York, NY: Portfolio, Sep. 2010.
-  X. Zhou, R. Li, T. Chen, and H. Zhang, “Network slicing as a service: Enable industries own software-defined cellular networks,” IEEE Commun. Mag., vol. 54, no. 7, pp. 146–153, Jul. 2016.
D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. v. d. Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis, “Mastering the game of Go with deep neural networks and tree search,”Nature, vol. 529, no. 7587, pp. 484–489, Jan. 2016. [Online]. Available: https://www.nature.com/articles/nature16961
-  W. Saad, M. Bennis, and M. Chen, “A vision of 6g wireless systems: Applications, trends, technologies, and open research problems,” arXiv:1902.10265 [cs, math], Feb. 2019, arXiv: 1902.10265. [Online]. Available: http://arxiv.org/abs/1902.10265
-  F. Tariq, M. Khandaker, K.-K. Wong, M. Imran, M. Bennis, and M. Debbah, “A speculative study on 6g,” arXiv:1902.06700 [cs], Feb. 2019, arXiv: 1902.06700. [Online]. Available: http://arxiv.org/abs/1902.06700
-  J. Hernández-Orallo and D. L. Dowe, “Measuring Universal Intelligence: Towards an Anytime Intelligence Test,” Artif. Intell., vol. 174, no. 18, pp. 1508–1539, Dec. 2010. [Online]. Available: http://dx.doi.org/10.1016/j.artint.2010.09.006
-  B. Kannan and L. E. Parker, “Metrics for quantifying system performance in intelligent, fault-tolerant multi-robot teams,” in Proc. IEEE IROS 2007, San Diego, CA, USA, Oct. 2007, pp. 951–958.
-  S. Legg and M. Hutter, “A formal measure of machine intelligence,” arXiv:cs/0605024, May 2006. [Online]. Available: http://arxiv.org/abs/cs/0605024
-  S. Liemhetcharat and M. Veloso, “Weighted synergy graphs for effective team formation with heterogeneous ad hoc agents,” Artif. Intell., vol. 208, pp. 41–65, Mar. 2014. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S0004370213001306
is now a assistant professor in Zhejiang University, Hangzhou, China. His research interests currently focus on multi-agent reinforcement learning and network slicing.
Zhifeng Zhao is an Associate Professor with Zhejiang University, China. His research area includes collective intelligence and software-defined networks.
Xing Xu is a Ph.D candidate in College of Computer Science and Technologies, Zhejiang University, Hangzhou, China. His research interests currently focus on collective intelligence.
Fei Ni is a Ph.D candidate in College of Computer Science and Technologies, Zhejiang University, Hangzhou, China. Her research interests currently focus on collective intelligence.
Honggang Zhang is a Professor of Zhejiang University, China. He is currently involved in the research on Cognitive Green Communications.