Increasing Trust in Meta-Heuristics by Using MAP-Elites Neil Urquhart School Of Computing, Edinburgh Napier University n.urquhart@napier.ac.uk Michael Guckert Technische Hochschule Mittelhessen michael.guckert@mnd.thm.de Simon Powers School Of Computing, Edinburgh Napier University s.powers@napier.ac.uk ABSTRACT Intelligent AI systems using approaches containing emergent el- ements often encounter acceptance problems. Results do not get sufciently explained and the procedure itself can not be fully retraced because the fow of control is dependent on stochastic ele- ments. Trust in such algorithms must be established so that users will accept results, without questioning whether the algorithm is sound. In this position paper we present an approach in which the user gets involved in the optimization procedure by letting them chose alternative solutions from a structure-archive which is cre- ated by the MAP-Elites algorithm. Analysis of these alternatives along the criteria of multiobjective optimization problems makes solutions comprehensible and hence is a means to build trust. We propose that the solution-focused nature of MAP-Elites allows the history of a solution to be easily shown to the user, explaining why that solution was included in those presented to the user. Here we demonstrate our ideas using a logistics problem previously explored by the authors. CCS CONCEPTS · Computing methodologies → Search methodologies; Dis- crete space search;· Applied computing → Transportation. KEYWORDS Vehicle Routing, Quality-Diversity Algorithms ACM Reference Format: Neil Urquhart, Michael Guckert, and Simon Powers. 2019. Increasing Trust in Meta-Heuristics by Using MAP-Elites. In Genetic and Evolutionary Com- putation Conference Companion (GECCO ’19 Companion), July 13ś17, 2019, Prague, Czech Republic. ACM, New York, NY, USA, 4 pages. https://doi.org/ 10.1145/3319619.3326816 1 INTRODUCTION 1.1 What is Trust? Complex intelligent systems that either contain elements in which they must frst be prepared with examples (e.g. supervised learning in neural networks) or contain stochastic elements (e.g. evolution- ary algorithms using stochastic operators) often meet acceptance problems as the way they compute their results is not transparent, Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specifc permission and/or a fee. Request permissions from permissions@acm.org. GECCO ’19 Companion, July 13ś17, 2019, Prague, Czech Republic © 2019 Association for Computing Machinery. ACM ISBN 978-1-4503-6748-6/19/07. . . $15.00 https://doi.org/10.1145/3319619.3326816 and a notion that they were achieved fortuitously persists for the non expert. For example, an evolutionary algorithm may search through millions of possible solutions, using operators such as re- combination and mutation to move between solutions. This can hardly be retraced without appropriate tools. Another example are algorithms that are trained on examples and therefore highly depend on the quality of the input data, and any bias within it. Because of elements that make use of emergent structures, repro- duction of decisions and retracing the algorithm ex post is difcult. Users applying such systems may be left in doubt about quality and validity of the results and are reluctant to accept them. Andras et al have outlined the idea that trust in intelligent machines can be achieved by repeated, successful use of a system (called inductive trust) [3]. This trust building process can be initiated by establish- ing mechanisms in which the results are explained and justifed in a form that is accessible to human beings. Various approaches to making such methods more explainable have been presented. Saving meta-information during the execution of the algorithm may facilitate the construction of an explanation of the fnal solu- tion. For example, neural networks may highlight data that was most infuential on the produced output and which was not [9]. Analogously, for population based approaches, an explanation that traces the evolution of the chosen solution may provide some form of justifcation for the fnal solution. 1.2 MAP-Elites The Multi-dimensional Archive of Phenotypic Elites (MAP-Elites) is an illumination algorithm that was frst introduced by Mouret et al [7]. MAP-Elites creates a structure archive of high-performing solutions mapped onto solution characteristics defned by the user. A set of characteristics is identifed which may be used to classify a solution (e.g. for a routing problem one might use cost, distance, delivery time span and vehicles required). Solutions are generated using mutation and recombination operators, but each solution can be classifed by normalizing its characteristics in order to identify a łbinž within the solution space that the solution belongs to. For instance we might normalize our four characteristics on a scale of 0- 20, thus a solution might occupy a bin such as 5:4:2:12 for example. The number of bins in a map is calculated as s d where s is the number of points on the scale and d is the number of dimensions. In our example the number of bins would be 160,000 (20 4 ). There exists the issue as to what happens when a solution is generated that belongs to a bin that is already occupied. In this case MAP- Elites uses a ftness value to determine which solution should be allowed to occupy the bin. In our vehicle routing problem (VRP) example we could utilize distance as the ftness value and thus, when a solution is found that maps to an occupied bin, it replaces the existing solution if it represents a decrease in distance. Figure 1