Capturing waste collection planning expert knowledge in a fitness function through preference learning

Until now, each local council has been requested waste collection services to the company Consorcio para la Gestión de los Residuos Sólidos de Asturias (COGERSA). The different requests have been appearing at different moments in time and every time a request has arisen, the experts of this company have just focused on a particular area to design specific routes, keeping unaltered the already planned routes. Hence, the actual routes have been locally optimized in each area object of request and one isolated from the rest. But now, the company requires unifying the waste collection planning in order to optimize their human and material resources, the collection time, fuel consumption and so on. The task is not as easy as it could seem because several issues must be taken into account. On one hand, there are several planning constraints imposed by the governments of each council. For instance, a container located at the entrance of the schools must not be picked up at the same time that the classes start or end. Also, the waste collection cannot be carried out when there is a street market planned. In addition, governments impose the collection frequencies and the number of containers. Also, the working time regulations must be considered, since, for instance, a same worker or truck cannot exceed a certain number of work hours. On the other hand, there are some other constraints about available resources. For instance, in regard to the laterality¹¹1Laterality is the property that has a container in order to indicate the side of the road to pick it up., the containers can be located to the right or to the left of the road, then, in case of two way roads, trucks must turn around in order to be able to pick up the container depending on the direction of traffic. Another issue to take into account is the distance that a truck travels with its cargo box loaded, since it might be better for the truck to pick up the containers on the way back to travel less distance with the cargo box full.
This problem is typically solved using planning optimization algorithms. The application of these techniques requires a fitness function able to evaluate the quality of a route planning. The main drawback is there exists a lack of a suitable function of this kind, and, even more, it is neither obvious nor easy to obtain in the waste collection. Hence, and firstly we need to face to the hurdle of designing a fitness function to evaluate a route planning, which will be the main focus and the main contribution of this paper. The plan in the future is to include this fitness function into an optimization algorithm scheme, but now this task is out of the scope of this paper. Despite the experts are not able to design a straightforward fitness function, they have a lot of expertise knowledge susceptible of being mined. Within artificial intelligence field, there exist two main ways of carrying out this practice. On one hand, if the experts are able to describe their knowledge, then this knowledge can be developed using rules leading to an expert system, for instance, for refrigerant flow air conditioning systems [13]. On the other hand, if only the actions of the experts are available but they are not able to explain them, then, it is possible to use machine learning techniques to extract their knowledge from those actions, for instance, for automated cyber security data triage [34]. The latter is the case of the COGERSA Company, since experts perform actions from their expertise, but they are not able to display the reasons about their decisions. There exist several ways of taking into account the expert knowledge to feed a machine learning approach. A straightforward possibility could be to ask the experts for providing a score for a route planning and then perform a learning process through a regression system. However, this kind of approaches are not adequate in general, since experts ratings cannot be interpreted as absolute assessments. The main reason is that experts tend to rate in a relative way, hence, the same expert might provide a different score to the same route planning depending on if this route planning is evaluated in a context with much more better route plannings or if it is evaluated in a context with much more worse route plannings. Thus, the route planning presented in a batch surrounded by worse route plannings will probably get a higher score than if this route planning were showed together with better route plannings. This situation is known as batch effect and it often biases the ratings. Hence, feeding a regression method with this unstable data can mislead the learning process [9]. However, it is possible to state that the expert assessments are consistent in regard to the order, that is, experts may provide the same order to the same route plannings set despite the score granted may differ from one expert to another. This evidence justifies that feeding a machine learning system with expert information in form of preference judgments is quite more promising than doing so in form of ratings, since avoiding the assumption that certain rating means the same for every expert avoids to mislead the learning process [5, 7]. This is the reason why the proposal of this paper advocates to fit the problem of finding an adequate fitness function to assess a route planning into a preference learning framework [3]. This environment not only requires to capture expert knowledge and expertise in form of preference judgements, otherwise the route plannings have to be described through indicators experts must establish in order to machine learning takes place. Those indicators will be the features of the machine learning process and they must capture the key issues experts implicitly pay attention to when they establish their preference judgements. The reality is that both kind of information were not available, hence, experts were asked for both preference judgments and key performance indicators. In fact, one of the contributions of this paper is an exhaustive design of key performance indicators to establish the route planning descriptions. Another contribution is to design route planning pairs from the actual route planning to make experts show their preferences, for which a greedy algorithm will be designed. The third contribution is the preference learning itself. Finally, experts reported a highly existing redundancy among the key indicators. At this respect, they manually proposed to take several sets to explore their impact. However, this paper goes further and the fourth contribution consists of carrying out an automatic feature selection procedure to establish the relevance of the key performance indicators according to a machine learning procedure. There are other approaches in the literature that provide a fitness function from key indicators. For instance, different key performance indicators from mobile networks [23] are combined using two approaches. The first one consists of a regression procedure, for which absolute ratings are required, a task almost unachievable by the experts. Besides, and, as commented before, this approach suffers from batch effect that can mislead the learning process. Even more, a regression procedure is not able to optimize the ranking [23]. The second approach optimizes the Spearman’s rank correlation using a particle swarm optimization, which is a population-based metaheuristic for solving continuous and discrete optimization problems. Our approach will be compared with this second approach in the experiments. Other works [29, 35] respectively proposes and uses in transport field a method called SWARA that combine key performance indicators. However, experts must sort them according to their significance. Also, relative significance must be stated beforehand, that is, it must be determined how much an indicator is more important than the next indicator in the sorting. At this respect, experts in the waste collection are not able to sort them, and must less to establish beforehand a relative significance, then, this method has been discarded. Also, another work [8] optimizes urban traffic flow using a genetic algorithm with Petri net analysis as fitness function. Their fitness function consists just in one key performance indicator, namely, the time, for whose estimation the Petri net is built. In our context, the time will also be a key performance indicator that will require to be estimated, but our proposal will involve performing a regression procedure. A Petri net cannot be applied in our context, since experts are required to build a Petri net. This task would involve establishing relationships among waste containers as the same way this work does among the roads using transitions. At this respect, experts do not report an existing relationship among containers. In fact, the original goal is to provide an optimized route planning, which means to establish the relationship among containers consisting in which container is collected next. They compare their approach with the lower distance. Hence, our approach will be compared with just consider the distance as fitness function. Another study [26] proposes a route planning and scheduling of waste collection and transport. The main drawback of this work is that it only uses one key performance indicator, namely, the time, which is estimated through a regression procedure from traveling distance. Our work goes further in the sense that experts reveal quite more issues involved in the route planning that just the time. For instance, they reveal the high influence of truck load and road kind as relevant aspects, so, this solution was discarded. All the same, our approach will be compared with the case of using the time key performance procedure as fitness function.

The rest of the paper is organized as follows. Section 2 describes some related work to this proposal. Section 3 details the whole process of building a promising fitness function for capturing the waste collection planning expert knowledge, including the design. The description of the experiments and the discussion of the results are exposed in Section 4. Finally, Section 5 draws some conclusions and proposes some lines of research for future work.

There are a lot of works that take the expert and human information to generate promising approaches that capture the expert knowledge. For example, a work [22] develops a model for detecting oil spill from expert information. Experts had to view images to detect suspicious regions and to classify them as positive or negative examples. Also, the experts provide the features, but the model obtained disregards many of them obtaining a good performance. Also, experts assess beef cattle as meat producers comparing animals with other partners [2]. Some body dimensions obtaining by a 3D photo describe the beefs. The expert knowledge was captured in a ranking function that unifies the preferences of different experts. Similarly, activity rubric is adjusted from the ranking that lecturers made from student answers [28]. The goal is to determine the weight of each activity item using partial rankings made by lectures from several student answers to the activity. Other study related to education [12] proposes a machine learning approach to obtain the most relevant factors that affect the employability and employment. To do that, some academics from different fields had to define the competencies that could be important for the employment using Spanish university data and student questionnaires. The aim of this study was to predict if a person will get employed or not and to extract the most relevant factors in order to know the best way for preparing the students for the labour market. Another work [19] presents a strategy to optimize the search engine retrieval quality. The user knowledge about the document ranking related to a query is given to the search engine through the click through data. The document the user clicks on is preferred with regard to the rest of documents. Support vector machines for ranking is fed with that information in order to generate a model able to rank the documents according to user preferences.

Before going in depth of obtaining a promising fitness function that captures the expert knowledge in waste collection planning, let us state some issues in order to better understand the current waste collection process. Actually, as commented before, there exists a set of routes locally planned by the experts from the successive council requests. The company has several garages in order to park the trucks. Each route begins in a garage and ends in the same garage. The company also has transfer stations located in different places of the region where the trucks can make one or more stopovers in order to unload the waste if the cargo box is too full during a route. Also, each route has a group of collection geographic points that can include more than one container (of the same or of different capacity). If there is more than one container at the same collection point, the capacity at that collection point will be considered as the accumulated capacity of all the containers at that point. A truck and a human team are assigned to each route. The human teams are formed by collection workers and by drivers. The number of workers of a team depends on the route and none worker can exceed 38 work hours per week. Also, workers must not work more than 8 hours per day and the trucks almost have a limit time of being driven.
A map of the container location together with the local designed routes allows concluding that the routes take the form of bouquets, since the majority have a central area with a high collection frequency and some branches with a low collection frequency. This is so because, the central area, which coincides with the settled area, is collected daily, but only one or two branches are included in the daily route. The experts have designed the routes per council (some routes cover more than one council if the areas are small) just using their experience and intuition causing the routes to have different duration. This is one of the pitfalls of the current waste collection planning, since there are some trucks and staff that overwork against others that underwork.
Some other issues include limitations in the container dimension in villages, the existence of vehicles that exceed the village weight and height limits or the allowed collection hour range at rush hours. Also, the possible delays in taking out the waste of some private companies because they have their container in a private enclosure or the adverse weather effects are other factors that make difficult the waste collection process. Besides, drivers might prefer to climb a road in worse conditions and to get down another road in better conditions than the opposite. Furthermore, the collection workers might prefer to collect the containers downhill than uphill. Another issue to keep in mind is the unloading time spent in transfer stations, which takes approximately two minutes. In addition to this, depending on the kind of truck, the running board, that is, the time to collect the waste can change. Moreover, the local governments determinate the amount of containers per surface and per population and decide the collection frequency of the container areas in terms of number of days. Owing to that, each container could have different frequency depending on the area. Furthermore, the collection frequencies of the containers may vary depending on the season of the year (Christmas time, summer, weekends…). It also happens that, for instance, in rural areas there are not time constraints to collect the waste and in some places container laterality is not an essential feature since sometimes the containers can be moved from one side of the road to the other. Other factor to take into account is the time that the driver takes from the garage to the first container on the route and from the last container to the garage.

Given the complexity and casuistry of the waste collection process taking into account all the above-mentioned issues, defining a fitness function for evaluating a whole waste collection planning in order to be used in the future in a planning optimization algorithm to produce a global waste collection planning is clearly a non trivial task. Our proposal is to capture the expertise and knowledge of the experts fitting the problem in a preference learning framework. In this context, one expects to have the disposal of waste collection planning pairs for which the experts must express their preferences. This means that a preference learning algorithm is fed by what is called preference judgments. A preference judgement is a pair $(x,y)$ of route plannings such that the route planning $x$ is preferred to the route planning $y$ because the quality of $x$, $f(x)$, is higher than the quality of $y$, $f(y)$, that is, $f(x)\geq f(y)$, but the values of $f(x)$ and $f(y)$ are unknown. Therefore, a preference learning algorithm is fed by the set

Then, the preference learning statement consists in obtaining a preference or ranking function $f$ defined in the route planning space such that $f(x)\geq f(y)$ whenever $x$ is preferred to $y$, that is, when $(x,y)\in S$. This means that $f$ establishes an order or rank for the route plannings involved in $S$, that is, applying this function to the route plannings in $S$ one can order them. However, this learning goes further, since $f$ is a promising discovery because it is able to yield a rating for whatever route planning $x$ out of $S$.

Now, it is necessary to describe a route planning $x$ through features and show to the experts pairs of route plannings from which they decide which member of the pair is preferable taking into account their expertise and knowledge. The idea is to obtain expertise and knowledge from the experts from two different points of view. The first one is to discover the key indicators experts consider relevant in order to obtain a route planning description in form of features. The second one is to find out the preference judgment of the experts looking at the route plannings themselves, without taking into account explicitly the key indicators. Then, the ranking or preference function $f$ is induced from these information sources in a attempt to reproduce the expert decision knowledge.

One of the difficulties that have arisen at this point is that the experts do not have the intellectual capacity of working in their minds with complete route plannings. They report serious difficulties in defining indicators for describing a whole route planning and also in deciding within a pair of route plannings which member of the pair is a better route planning. Hence, our proposal also includes making affordable the work of the experts in this sense. The idea at this respect consists of reducing the initial problem from the route planning space to the route space. In fact, up to now experts get use of designing routes in local areas, one isolated from the rest, so this simplification clearly makes the task more affordable for them. Then, pairs of routes rather than pairs of route plannings are presented to the experts, so they must decide which route is better in each pair of routes rather than decide which planning is better in each pair of route plannings. Also, our proposal establishes several key performance indicators to describe a route rather than a route planning taking into account the above-mentioned issues and factors about the waste collection process and the information coming from the experts. The indicators will be strategically designed in order to satisfy the additivity property. This feature makes them be suitable to further extend the solution from the route space to the route planning space.

The discussion of the key performance indicators established is detailed in Section 3.1. One drawback arisen in designing these indicators is that travel time is not an available data, despite the experts state that it is a key issue to take into account. This is one of the obstacles that have been arisen in the design of key performance indicators. This work overcomes this drawback proposing a regression procedure in order to estimate the travel time distance from other available information. Up to now, only current local routes designed by the experts are available, but including the problem in a preference learning environment requires to have the disposal of other routes to built pairs from which experts must decide the member of each pair they prefer. In this sense, it arises the necessity of creating alternative routes able to design adequate route pairs to successfully allow acquiring the expert knowledge. This is another obstacle that has been arisen in the research of this paper. Our proposal copes with this drawback designing a greedy algorithm to create alternative routes for this purpose. The details are explained in Section 3.2. Finally, building a fitness function for evaluating a route through a preference learning process is exposed in Section 3.3.

This section deals with the experiments carried out in this work, but before establishing the parameter settings and displaying and discussing the experiment results, let us state the geographic information systems for obtaining the KPIs (see Section 4.1). Concerning the experiments, two different types are exposed, namely, i) experiments to estimate the travel time KPI through a regression process (see Section 4.2) and ii) experiments through a preference learning process to provide the fitness function to evaluate the quality of a waste collection route that, as it was previously mentioned, it will be also taken to evaluate the quality of a waste collection planning (see Section 4.3); this last kind of experiments also includes the feature selection process.

This paper takes the first step in establishing an automatic and global optimization on the waste collection process of the company called Consorcio para la Gestión de los Residuos Sólidos de Asturias (COGERSA). This step consists of designing a fitness function to assess a candidate route planning in order to use it in planning optimization algorithm for providing a global optimal route planning. So far, the experts manually designed the route planning focusing on particular local areas with a relative reduced number of waste containers whenever and under council demand. The route planning complexity due to the existing resources and constraints even makes the task of defining a fitness function not be affordable by the experts. This paper provides a fitness function from the expert knowledge and stating its design in a preference learning framework. The process is simplified to design a fitness function that assesses a route rather than a route planning again due to complexity. Fortunately, this simplification was easy to carry out, since expert did not mention specific indicators of a planning route, maybe because so far they planned routes one isolated from the rest. The idea under preference learning is to obtain expertise and knowledge from the experts from two different points of view. The first one is to discover the key indicators experts consider relevant in order to obtain a route description in form of features. The second one is to find out the preference judgment of the experts looking at the routes themselves, without taking into account explicitly the key indicators. Then, a ranking or preference function is induced from these information sources in an attempt to reproduce the expert decision knowledge, which will be the fitness function. For this purpose, this paper has exhaustively and carefully designed several key performance indicators according to the experts. These indicators were established for describing a route, but all of them have the additivity property, hence, they are straightforward extended to a route planning. Also, experts expect redundancy among such indicators, and for this reason, a feature selection analysis was included in this paper in order to check in what extent this hypothesis holds. The conclusion of the study confirms the existence of high redundancy among the indicators, since the best performance is reached when $6$ or $8$ out of $21$ indicators are taken. Truck load seems to be a clue indicator together with the distance travelled, especially if such distance belongs to no main roads. The proposed approach has been compared to other existing methods in the literature, that just consider a unique indicator as fitness function or combine them through a particle swarm optimization. The conclusion is that the method proposed in this paper clearly outperforms the existing approaches, since the performance goes from $72\%$ for a unique indicator fitness function or $90\%$ for particle swarm optimization to $98\%$ of the proposed preference learning approach.
As future work, the planning consists of including the fitness function in planning optimization algorithms. Besides, and given the existing redundancy among the key performance indicators, we plan to study more in depth those that have been shown to be clue in this study. For instance, truck load is one of the most relevant indicators. However, the way experts have found for computing it is quite gross. Hence, obtaining an alternative way of estimating it in a more accurate way may lead to an improvement. Also, the distance travelled along different kinds of roads seems to determine the quality of the route. In this sense, it may be promising to polish the split of the roads in different kinds, taking into account not only the official type of the route, otherwise the state, traffic and other environment issues. Finally, and since the original goal was to optimize a route planning it would be the great interest to consider key performance indicators of route planning, for instance, the number of routes, highly related with the human and material resources.

This research has been partially supported by the Spanish Ministerio de Ciencia e Innovación through the grant PID2019-110742RB-I00) and by the RDI project Smart Waste Collection (SWC), developed by the consortium COGERSA, SADIM S.A., S.M.E - grupo HUNOSA and ABAMobile and partially supported by the Instituto de Desarrollo Económico del Principado de Asturias (IDEPA) through the grants IDE/2015/000863, IDE/2015/000864 and IDE/2015/000865.