1、agent controlled traffic lightsintroductionthe quality of (urban) traffic control systems is determined by the match between the control schema and the actual traffic patterns. if traffic patterns change, what they usually do, the effectiveness is determined by the way in which the system adapts to

2、these changes. when this ability to adapt becomes an integral part of the traffic control unit it can react better to changes in traffic conditions. adjusting a traffic control unit is a costly and timely affair if it involves human attention. the hypothesis is that it might offer additional benefit

3、 using self-evaluating and self-adjusting traffic control systems. there is already a market for an urban traffic control system that is able to react if the environment changes;the so called adaptive systems. real adaptive systems will need pro-active calculated traffic information and cycle plans-

4、 based on these calculated traffic conditions- to be updated frequently.adaptive urban traffic control adaptive signal control systems must have a capability to optimise the traffic flow by adjusting the traffic signals based on current traffic. all used traffic signal control methods are based on f

5、eed-back algorithms using traffic demand data -varying from years to a couple of minutes - in the past. current adaptive systems often operate on the basis of adaptive green phases and flexible co-ordination in (sub)networks based on measured traffic conditions (e.g., utopia-spot,scoot). these metho

6、ds are still not optimal where traffic demand changes rapidly within a short time interval. the basic premise is that existing signal plan generation tools make rational decisions about signal plans under varying conditions; but almost none of the current available tools behave pro-actively or have

7、meta-rules that may change behaviour of the controller incorporated into the system. the next logical step for traffic control is the inclusion of these meta-rules and pro active and goal-oriented behaviour. the key aspects of improved control, for which contributions from artificial intelligence an

8、d artificial intelligent agents can be expected, include the capability of dealing with conflicting objectives; the capability of making pro-active decisions on the basis of temporal analysis; the ability of managing, learning, self adjusting and responding to non-recurrent and unexpected events (am

9、brosino et al., 1994).what are intelligent agents agent technology is a new concept within the artificial intelligence (ai). the agent paradigm in ai is based upon the notion of reactive, autonomous, internally-motivated entities that inhabit dynamic, not necessarily fully predictable environments (

10、weiss, 1999). autonomy is the ability to function as an independent unit over an extended period of time, performing a variety of actions necessary to achieve pre-designated objectives while responding to stimuli produced by integrally contained sensors (ziegler, 1990). multi-agent systems can be ch

11、aracterised by the interaction of many agents trying to solve a variety of problems in a co-operative fashion. besides ai, intelligent agents should have some additional attributes to solve problems by itself in real-time; understand information; have goals and intentions; draw distinctions between

12、situations; generalise; synthesise new concepts and / or ideas; model the world they operate in and plan and predict consequences of actions and evaluate alternatives. the problem solving component of an intelligent agent can be a rule-based system but can also be a neural network or a fuzzy expert

13、system. it may be obvious that finding a feasible solution is a necessity for an agent. often local optima in decentralised systems, are not the global optimum. this problem is not easily solved. the solution has to be found by tailoring the interaction mechanism or to have a supervising agent co-or

14、dinating the optimisation process of the other agents.intelligent agents in utc,a helpful paradigmagent technology is applicable in different fields within utc. the ones most important mentioning are: information agents, agents for traffic simulation and traffic control. currently, most applications

15、 of intelligent agents are information agents. they collect information via a network. with special designed agents user specific information can be provided. in urban traffic these intelligent agents are useable in delivering information about weather, traffic jams, public transport, route closures

16、, best routes, etc. to the user via a personal travel assistant. agent technology can also be used for aggregating data for further distribution. agents and multi agent systems are capable of simulating complex systems for traffic simulation. these systems often use one agent for every traffic parti

17、cipant (in a similar way as object oriented programs often use objects). the application of agents in (urban) traffic control is the one that has our prime interest. here we ultimately want to use agents for pro-active traffic light control with on-line optimisation. signal plans then will be determ

18、ined based on predicted and measured detector data and will be tuned with adjoining agents. the most promising aspects of agent technology, the flexibility and pro-active behaviour, give utc the possibility of better anticipation of traffic. current utc is not that flexible, it is unable to adjust i

19、tself if situations change and cant handle un-programmed situations. agent technology can also be implemented on several different control layers. this gives the advantage of being close to current utc while leaving considerable freedom at the lower (intersection) level.assumptions and consideration

20、s on agent based urban traffic controlthere are three aspects where agent based traffic control and -management can improve current state of the art utc systems:- adaptability. intelligent agents are able to adapt its behaviour and can learn from earlier situations.- communication. communication mak

21、es it possible for agents to co-operate and tune signal plans.- pro-active behaviour. due to the pro active behaviour traffic control systems are able to plan ahead.to be acceptable as replacement unit for current traffic control units, the system should perform the same or better than current syste

22、ms. the agent based utc will require on-line and pro-active reaction on changing traffic patterns. an agent based utc should be demand responsive as well as adaptive during all stages and times. new methods for traffic control and traffic prediction should be developed as current ones do not suffice

23、 and cannot be used in agent technology. the adaptability can also be divided in several different time scales where the system may need to handle in a different way (rogier, 1999):- gradual changes due to changing traffic volumes over a longer period of time,- abrupt changes due to changing traffic

24、 volumes over a longer period of time,- abrupt, temporal, changes due to changing traffic volumes over a short period of time,- abrupt, temporal, changes due to prioritised traffic over a short period of timeone way of handling the balance between performance and complexity is the use of a hierarchi

25、cal system layout. we propose a hierarchy of agents where every agent is responsible for its own optimal solution, but may not only be influenced by adjoining agents but also via higher level agents. these agents have the task of solving conflicts between lower level agents that they cant solve. thi

26、s represents current traffic control implementations and ideas. one final aspect to be mentioned is the robustness of agent based systems (if all communication fails the agent runs on, if the agent fails a fixed program can be executed.to be able to keep our first urban traffic control model as simp

27、le as possible we have made the following assumptions: we limit ourselves to inner city traffic control (road segments, intersections, corridors), we handle only controlled intersections with detectors (intensity and speed) at all road segments, we only handle cars and we use simple rule bases for k

28、nowledge representation.types of agents in urban intersection control as we divide the system in several, recognisable, parts we define the following 4 types of agents:- roads are represented by special road segment agents (rsa),- controlled intersections are represented by intersection agents (itsa

29、),- for specific, defined, areas there is an area agent (higher level),- for specific routes there can be route agents, that spans several adjoining road segments (higher level).we have not chosen for one agent per signal. this may result in a more simple solution but available traffic control progr

30、ams do not fit in that kind of agent. we deliberately choose a more complex agent to be able to use standard traffic control design algorithms and programs. the idea still is the optimisation on a local level (intersection), but with local and global control. therefor we use area agents and route ag

31、ents. all communication takes place between neighbouring agents and upper and lower level ones.design of our agent based systemthe essence of a, demand responsive and pro-active agent based utc consists of several itsas (intersection agent).,some authority agents (area and route agents) and optional

32、 road segment agents (rsa). the itsa makes decisions on how to control its intersection based on its goals, capability, knowledge, perception and data. when necessary an agent can request for additional information or receive other goals or orders from its authority agent(s).for a specific itsa, imp

33、lemented to serve as an urban traffic control agent, the following actions are incorporated (roozemond, 1998):- data collection / distribution (via rsa - information on the current state of traffic; from / to other itsas - on other adjoining signalised intersections);- analysis (with an accurate mod

34、el of the surrounds and knowing the traffic and traffic control rules define current trend; detect current traffic problems);- calculation (calculate the next, optimal, cycle mathematically correct);- decision making (with other agent deciding what to use for next cycle; handle current traffic probl

35、ems);- control (operate the signals according to cycle plan).the internals of the itsa modeltraffic dependent intersection control normally works in a fast loop. the detector data is fed into the control algorithm. based upon predetermined rules a control strategy is chosen and the signals are opera

36、ted accordingly. in this research we suggest the introduction of an extra, slow, loop where rules and parameters of a prediction- model can be changed by a higher order meta-model.itsa modelthe internals of an itsa consists of several agents. for a better overview of the internal itsa model-agents a

37、nd agent based functions see figure 2. data collection is partly placed at the rsas and partly placed in the itsas. the needed data is collected from different sources, but mainly via detectors. the data is stored locally and may be transmitted to other agents. the actual operation of the traffic si

38、gnals is left to an itsa-controller agent. the central part of the itsa, acts as a control strategy agent. that agent can operate several control strategies, such as anti-blocking and public transport priority strategies. the control strategy agent uses the estimates of the prediction model agent wh

39、ich estimates the states in the near future. the itsa-prediction model agent estimates the states in the near future. the prediction model agent gets its data related to intersection and road segments - as an agent that knows the forecasting equations, actual traffic conditions and constraints - and

40、 future traffic situations can be calculated by way of an inference engine and its knowledge and data base. on-line optimisation only works if there is sufficient quality in traffic predictions, a good choice is made regarding the performance indicators and an effective way is found to handle one-ti

41、me occurrences (rogier, 1999).prediction modelwe hope to include pro-activeness via specific prediction model agents with a task of predicting future traffic conditions. the prediction models are extremely important for the development of pro active traffic control. the proposed itsa-prediction mode

42、l agent estimates the states of the traffic in the near future via its own prediction model. the prediction meta-model compares the accuracy of the predictions with current traffic and will adjust the prediction parameters if the predictions were insufficient or not accurate. the prediction model ag

43、ent is fed by several inputs: vehicle detection system, relevant road conditions, control strategies, important data on this intersection and its traffic condition, communication with itsas of nearby intersections and higher level agents. the agent itself has a rule-base, forecasting equations, know

44、s constraints regarding specific intersections and gets insight into current (traffic) conditions. with these data future traffic situations should be calculated by its internal traffic forecasting model. the predicted forecast is valid for a limited time. research has shown that models using histor

45、ic, up-stream and current link traffic give the best results (hobeika & kim, 1994).conclusions and future workadaptive signal control systems that are able to optimise and adjust the signal settings are able to improve the vehicular throughput and minimise delay through appropriate response to chang

46、es in the measured demand patterns. with the introduction of two un-coupled feed back loops, whether agent technology is used or not, a pro-active theory of traffic control can be met. there are several aspects still unresearched. the first thing we are going to do is to build a prototype system of

47、a single intersection to see if the given claims of adaptability and pro activeness can be realised. a working prototype of such system should give appropriate evidence on the usability of agent based control systems. there are three other major subjects to be researched in depth; namely self adjust

48、able control schemas, on-line optimisation of complex systems and getting good prediction models. for urban traffic control we need to develop self adjustable control schemes that can deal with dynamic and actuated data. for the optimisation we need mathematical programming methodologies capable of

49、real-time on-line operation. in arterial and agent based systems this subject becomes complex due to several different, continuously changing, weights and different goals of the different itsas and due to the need for co-ordination and synchronisation. the research towards realising real-time on-lin

50、e prediction models needs to be developed in compliance with agent based technology. the pro-active and re-active nature of agents and the double loop control schema seems to be a helpful paradigm in intelligent traffic management and control. further research and simulated tests on a control strate

51、gy, based on intelligent autonomous agents, is necessary to provide appropriate evidence on the usability of agent-based control systems. 代理控制交通燈前言（城市）交通控制系統的好壞決定于系統控制模式和實際交通流量模式是否相符。如果交通方式改變，他們通常所采取的有效措施是使系統適應這些改變。當這種改變能力成為交通控制系統單元的一部分，它將能更好地改變目前的交通狀況。如果考慮人工調整，那么調整交通控制單元是一件花費昂貴和耗時大的事情。假如使用一種能夠自我分析和

52、自適應調整的交通控制系統，它將會帶來額外的收益。對于這種可以適應環境改變的城市交通控制系統已經有銷售。這種系統叫做自適應系統。“實時”自適應系統將需要主動計算交通信息和循環地計劃，這種計劃是根據不斷計算更新的交通信息而采取的。自適應城市交通控制自適應信號控制系統必須有能力根據目前的交通狀況，通過調整優化交通流量。所有使用的交通信號控制方法是基于反饋算法，而反饋算法的數據來源于過去的幾分鐘到幾年時間里的交通需求。當前的交通控制系統運作網絡是根據測到的交通狀況來協調適應性和靈活性。這些還不是根據交通需求在短時間內改變的最佳方法。其基本前提是，現有的信號計劃工具能夠根據不同的情況作出合適的信號計劃。

53、但是，目前可用的工具都是積極主動型，或者元規則型。這就使得改變控制器的方式放在系統中進行。交通控制的下一個邏輯步驟是將元規則、主動的和目標導向行為都列入。為此，從人工智能和人工智能代理的貢獻可以預測，加強管制的主要方面包括：與沖突的目標處理能力；基于時間分析的決策能力；管理、學習、自我調整和應對非經常性和突發事件的能力（ambrosino等.，1994）。什么是智能代理 代理技術是人工智能的一個新概念。人工智能代理模式是基于無功，自主，內部機動的實體，動態的居住概念，并不一定是完全可預見的環境（weiss，1999年）。自主是指在一段長時間內，能夠作為一個獨立單元，執行必要的行動以實現多種預先

54、指定的目標，同時對同一體載傳感器產生的刺激產生反應（ziegler, 1990）。綜合代理系統的特點是可以通過很多的代理系統互相合作去解決各種各樣的問題。此外，人工智能、智能代理應該有一些額外的屬性使得它自己可以解決一些實時問題、理解信息、有目標和方向、對情況進行分清區別、揉合新概念或想法。其解決問題的智能代理組件可以是一個以規則為基礎的系統，但也可以是一個神經網絡或模糊專家系統。顯然，對代理技術而言，找到一個可行的解決方案是必要的。通常在分散系統中最優的是局部而不是全部。這個問題不容易解決。解決的方案可以參照裁縫業的相互作用機制，或者有監督代理人協調其他代理人進行優化的過程。智能代理在utc

55、的使用是一種有益的范例代理技術適合不同領域的utc。那些值得最重要一提的是：信息代理、代理交通仿真和交通控制。目前，大多數運用智能代理來進行信息代理。它們通過網絡收集信息。對于采用特殊設計的代理商用戶可以提供具體的資料。城市交通智能代理可以為用戶提供有關的天氣信息，交通擁堵，公共交通，道路封鎖，最佳路線等，成為個人的旅行助理。代理技術也可以用于數據的分配。代理和多代理系統，有模擬仿真復雜交通系統的能力。這些系統通常為每一個交通參與者使用一個代理（類似于面向對象程序使用對象）。我們感興趣的是代理技術在（城市）交通控制中的應用。為此，我們最終使用實時代理技術進行有效的交通燈控制。信號計劃將根據預測

56、和探測器實測數據，以及相鄰的代理控制的基礎上而進行調整。代理技術最具有優勢的是它的靈活性和積極性，可以給utc帶來更好的交通預測。當前的utc不具有靈活性的，它無法自行調節，如果情況變化，不能處理非程序化的情況。代理技術可以在不同的控制層進行控制。這對于接近目前還處于低水平的utc具有一定的優勢。基于城市交通控制的代理假設和考慮基于城市交通控制和管理的代理，可以從三方面改善utc系統的現狀。l 適應性。智能代理能夠適應它的行為，可以學習早期的情況。l 通信。通信是代理之間進行合作，從而對計劃信號進行調整。l 積極主動。由于積極主動，交通控制控制系統可以提前計劃。要對目前的交通控制單元進行替換，

57、可以接受的是該系統應執行得和現在系統一樣或者比它更好。基于utc的代理將要求改變交通模式的在線反應和積極反應。基于utc的代理應根據各個階段的要求做出適應性的回應, 當前的交通控制和預測方法是不夠的，不能使用在代理技術上，應該發展新的交通控制和交通預測方法。適應性也可以分為幾個不同的時間尺度，系統可能需要以不同的方式處理（rogier，1999年）：l 由于漸進式的改革改變了一個較長時間的交通量；lll .由于突然的變化改變了一個較長時間的交通量；l 突然，由于改用短的時間內的交通量，時空改變；l 突然，時空改變，由于交通優先較短時間。一種處理方法之間的性能和復雜性是一個等級制度布局使用的平衡。我們提出了一個在每一個層次的代理對自己負責的最佳解決方案，可能不僅影響鄰近的代理，而且也可能影響更高一級代理。這些代理的下級代理之間解決他們不能解決的沖突任務。最后要提的一個方面是穩健的代理系統（如果所有運行的代理通訊失敗，如果沒有一個固定的代理程序被執行）。為了能夠保持我們的第一個城市交通控制模型盡可能簡單，我們作了以下的假設：我們限制自己的市內交通控制（路段，路口，走廊），我們只處理控制與探測器（強度和速度）在所有路段的十字路口，只處理車和我們使用簡單的規則為基礎