In this paper we sketch a vision of computable law as argumentation-based MAS, i.e., human-centred intelligent systems densely populated by agents (software or human) capable of understanding, arguing, and reporting, via factual assertions and arguments, about what is happening and what they can make possibly happen. A multi-agent system based on argumentation, dialogue, and conversation is, in this vision, the basis for making the law computable: through argumentation, dialogue, and adherence to social judgment, the behaviour of the intelligent system can be reached, shaped, and controlled with respect to the law. In such a scenario, computable law - and related intelligent behaviour - is likely to become associated with the capability of arguing about state and situation, by reaching a consensus on what is happening around and what is needed, and by triggering and orchestrating proper decentralised semantic conversations to decide how to collectively act in order to reach a future desirable state. Interpretability and explainability become important features for that sort of systems, based on the integration of logic-based and sub-symbolic techniques. Within this novel setting, MAS methodologies and technologies become the starting point to achieve computable law, even if they need to be adapted and extended for dealing with new challenges. Accordingly, in this paper we discuss how this novel vision can build upon some readily-available technologies, and the research challenges it poses. We analyse a number of approaches and technologies that should be involved in the engineering of systems and services, and become core expertise for distributed systems engineers. Among the others, these include knowledge representation, machine learning, and logic argumentation.
The research field of computable law studies the engineering of the law – i.e., the design of
appropriate formal models and the development of the corresponding technology – to allow
norms, terms and conditions to be represented in a machine-understandable way [
The second factor is that nowadays application scenarios for computable law are strongly characterised by the same “symbolic vs sub-symbolic” dichotomy that also informs the most promising techniques in artificial intelligence today. Indeed, if, on the one hand, legal intelligence has historically been bound to logic and argumentation (i.e., symbolic approaches), on the other hand, the applications of machine learning algorithms (i.e., sub-symbolic approaches) are increasing today, especially in the field of data analysis and predictive justice. The use of the latter techniques raises ethical and fairness issues linked to the lack of transparency and the possibility of hidden biases. Therefore, models aimed at making the law computable must somehow consider that dichotomy and should try a reconciliation – possibly via a blended integration – so to take advantage of each approach: benefiting from the strengths of each method, and smoothing the corresponding limits.
Along this line, in this paper we sketch a vision of computable law as an argumentation-based
multi-agent system. As widely recognised, agent architecture is today the reference for the
design of intelligent systems [
In this scenario, the very nature of the system actors – intended as agents, but also as surrounding environments – embodies the concepts of computable law implementing and coordinating the activities of distributed processes in order to achieve either individual or common goals. In fact, computable law is likely to become associated with the capability of debating about situations and about the current context, by reaching a consensus on what is happening around and what is needed, and by triggering and directing proper decentralised semantic conversations to decide how to collectively act in order to reach the future desirable state of the afairs. Within this novel setting, interpretability and explainability become a remarkable feature of the multi-agent system.
Based on the envisioned scenario, the contributions of this paper are the following: • We detail the concept of computable law as argumentation-based MAS (i.e., conversational agents), also with the help of a running example, and show how they afect the engineering of intelligent systems, challenging traditional approaches to distributed computing and calling for novel argumentation approaches. • We investigate a number of approaches and technologies that should be involved in the engineering of systems and services in that original scenario, and should become core expertise for distributed systems engineering. Among the others, those include knowledge representation and ontologies, machine learning, argumentation models and technologies, human-computer interfaces.
To ground the discussion, we first examine a case study in the area of trafic management, considering the near future of self-driving cars. In that scenario cars are capable of communicating with each other and with the road infrastructure while cities and roads are suitably enriched with sensors and virtual trafic signs able to dynamically interact with cars to provide information and supervision.
The choice of the case study is driven by the fact that both self-driving cars and the whole intelligent transportation domain – including trafic management – are a natural fit for computable law since they need real-time control and feedback – with respect to the norms and the current state of afairs – and should both adapt to legislation as well as to possible contingencies (also involving ethical choices and legal reasoning).
Accordingly, self-driving cars need to (i) exhibit some degree of intelligence for taking autonomous decisions, (ii) converse with the context that surrounds them, (iii) have humans in the loop, and (iv) be deeply intertwined with the law setting characterising the environment and the society. Generally speaking, the main success factor to address the goal of the car (i.e., reach a destination) is the capability to converse and dialectically interact with the surrounding environment (other vehicles, infrastructure, humans, etc.), in order to make the best choice and therefore actuate the best action.
Part of this scenario is already a reality in many (smart) cities around the world. There, IT
technologies are exploited to improve both the organisation of the transportation infrastructure
and the performance of some specific parts. Adaptive trafic lights, dynamic speed and flow
metering, urban monitoring stations, and similar tools and algorithms are routinely employed by
district administrations as a means to observe and revise trafic situations in (almost) real-time
[
In the following we analyse the scenario from a computational perspective – i.e., discussing in details actors and actions that come up as well as system requirements – in order to provide the bases for its reification on the most appropriate engineering approaches and corresponding architecture.
In the envisioned desiderata scenario, passengers simply express their desire (e.g., “bring me to the hospital") and the car starts acting in autonomy, travelling towards the destination and without passengers to worry about the specific actions and decisions to be undertaken. There, autonomous goal-oriented smart agents (either software, objects, humans, etc.) are pervasive since multiple actors come into play, such as other self-driving cars or, for instance, autonomous intersection managers that regulate the flow of cars based on specific goals imposed by the municipality (e.g., reduce circulation in a specific area). Moreover, the environment is one of the main actors of such a system–providing context information, such as rules valid in that specific area or situation, i.e., encapsulating context knowledge and intelligence.
Looking in-depth at the expect behaviour of self-driving cars in an urban environment, it turns out that autonomous cars need to undertake a complex decision-making process, in compliance with norms and social rules, in almost the whole journey, which seamlessly integrates actions (and perceptions) at diferent levels.
The first action is the trip planning which must be done by considering the global knowledge— for instance, local laws, regulations, policies, and average trafic conditions of the intersections and roads along the path toward the destination, influenced by factors such as the day of the week, the hour of the day, the weather, and the like.
Planning is then modulated by considering all contingencies arising during the journey – such as, for instance, a car crash forcing a change path, a protest causing delays in our preferred path, etc. As a consequence, the original plan has to be adapted to the local knowledge (and perceptions) cars gather while enacting it; once again, the decision is closely related to legal issues: for instance, what is the best action to perform in order to avoid fines? Or, again, if an accident cannot be avoided, how to choose the ethically-preferable option?
Moreover, there are strict rules on safety we would like self-driving cars to automatically enforce, such as slowing down when passing nearby a school, breaking and setting aside the vehicle as soon as the horn of an ambulance is heard, etc. In other words, regardless of what the global and local knowledge may suggest, we abide by a set of general commonsense rules orthogonally considered valid.
Generally speaking, compliance with the legal rules by the autonomous cars – including also global, national, state, and local laws, regulations, and policies – must be guaranteed, unless contingent priority situations occur (such as when a norm has to be violated in order to save the life of a passenger/pedestrian).
The analysis carried out so far highlights some key requirements that engineering approaches and techniques have to fulfil. First of all, the envisioned scenario seamlessly integrates perceptions (and actions) at two diferent scales—namely, the macro and the micro.
The macro level of the system includes global knowledge and generally-valid rules, like universal norms and legal conventions, possibly modulated by commonsense reasoning. Moreover, with respect to the surrounding environment, the macro level deals with a mid / long term horizon and focus on the issue of trafic flow management–including, for instance, trafic flow forecasting and urban planning possibly learned from historical data analysis. On the other hand, the micro level, deals with the short term horizon, and mostly focuses on intersection management, including a few highly intertwined sub-problems—e.g. collision avoidance, minimisation of average delay, and congestion resolution. The macro-level and the micro-level act synergistically, exploiting some sort of integration in order to achieve individual and social goals.
As the last step, a suitable infrastructure for V2I (vehicle-to-infrastructure) and V2V (vehicle-tovehicle) communication should be considered, for instance through the deployment of Road-side Units (RSU). This infrastructure should make it possible to convey information from the vehicle to the infrastructure and vice-versa. For instance, it should provide information about the road on which the vehicle is travelling and its specific rules, environmental conditions around the vehicle, trafic in the vicinity of the vehicle, and construction in the vicinity of the vehicle.
All the abovementioned ingredients should be mixed consistently in order to achieve the goal of the cars and therefore of the user, respecting current rules and convention. Therefore, the system knowledge base needs to be built by taking into account both macro and micro scales, and agents have to be able to reason and argue over it in order to achieve their goals.
In the following, we show that computable law naturally is an essential ingredient in a distributed multi-party conversation, or dialogue, based on distributed intelligence. It cannot be easily tackled with traditional approaches to distributed computing: instead, diferent approaches and techniques needs to be put in place and to be fruitfully integrated in order to meet the aforementioned requirements.
The wide variety of actors and requirements in the above-described scenario recall the MAS
model and architecture as intrinsically suitable for addressing issues and challenges lay ahead.
Along this line, the vision of computable law described in this paper is based on the fundamental
roles of MAS – individuals (agents), society, and environment [
Then, our vision enhances these fundamentals roles taking into account two more key concepts and related abstraction: norms and e-institutions – i.e., proper roles and abstractions for the normative environment – and micro-intelligence—i.e., rational reasoning capabilities to enhance both environment and individuals.
In a nutshell, the system is composed of several agents, each with his own personal goal to achieve. Agents’ interaction and dialogue, on normative aspects and surrounding situations, allow them to complete their goals. For that reason, rules and conventions must have an active role in the system (e-institution) in order to be questioned, examined and respected. The system’s actors – individuals, environment, e-institution – are intelligent (micro-intelligence) in that they can reason on knowledge and context, argue and explain the rationale behind their decision. The overall behaviour of the system transposes the social behaviour, the debate with the society can act as a social judgment on individuals and therefore can calibrate their individual actions.
Diferent roles of e-institution and micro-intelligence and their interaction are discussed below.
In the vision sketched above, particular attention has been given to normative concepts since
norms and laws are the basic bricks upon which to build the concept of computable law, since
individual and collective behaviours are both afected by norms. In particular, we leverage
on the well-known concepts of electronic-institution (e-institution) – as in normative MAS
[
Loosely speaking, e-institutions are computational realisations of traditional institutions:
i.e., coordination artefacts providing an environment where agents can interact according to
stated laws – norms or conventions – in such a way that interactions within the e-institution
reproduce norm-based interactions in the actual world. With the term deliberative agents, we
emphasise the agents’ autonomy stressed both by e-institutions and normative systems. Indeed,
in such systems individuals possess the property of normative autonomy—i.e., can decide to
violate a norm to achieve their goals, or to change their goals so that they match the existing
norms. E-instutions can provide for real-time detection of violations, and norm-enforcement
can be envisaged via diferent enforcement technique. For instance, a simple local blocking rule
[
From an agent’s point of view, two major benefits stem from modeling environments as e-institutions. On the one hand, e-institutions establish conventions – on behaviour, language, and protocols – that inform agents about the law and possibly induce them to comply. In a sense, the environment is given structure, so that the agents have an easy comprehension of its working laws. On the other hand, norm enforcer agents endowed with capabilities for acquiring norms dynamically and enforcing them in uncertain environments can be envisioned and spread all over the system. Recalling our motivating scenario: drivers or trafic wardens know when they can talk, what consequences their acts will have, and what actions are possible at each moment in time. These restrictions contains the set of actions that agents have to consider at each moment in time by limiting the set of options that agents have to think about.
From the systems properties’ point of view, e-institution promote a clear embodiment of the laws that govern the system, thus making it observable and more explainable in its autonomous actions—i.e., the explicit role of institution allow to enforce a set of norms whose violation can be perfectly observed. With respect to the case study, e-institution incarnates global system knowledge as global, national, state, and local laws, regulations, and policies.
In order to ensure diferent levels of knowledge – and intelligence – as highlighted in Section 2,
the MAS model is extended with the concept of micro-intelligence. Micro-intelligence is exploited
to manage precisely the micro-level of the system intelligence – i.e., local intelligence, as the
intelligence of objects, things, but also of the environment – to be integrated with the
macrolevel intelligence. We recall the micro-intelligence definition from [
Operationally, micro-intelligence is about scattering small chunks of machine intelligence all over a distributed and situated system, capable to enable the individual intelligence and the argumentation capability of any sort of devices. Micro-intelligence can be encapsulated in devices of any sort, making them smart, and capable to work together in groups, aggregates, societies. Thus, the micro-intelligence vision promotes ubiquitous distribution of intelligence in large pervasive systems such as those belonging to the pervasive and IoT landscape, in particular as a complement to agent-based technologies and methods, at both the individual and the collective level.
Note the micro-intelligence model – and related technologies – becomes fundamental also for the e-institution since it provides externalised rationality for reasoning and speaking with other system actors as well as the capability to reach properties such as interpretability and explainability, by leveraging on symbolic approach.
Moreover, micro-intelligence, as depicted in this summary, becomes the fundamental engine enabling conversation and argumentation between the system’s actors and facilitating the control of contingent situations establishing the possibility to reach a consensus among speaking agents.
Mixing up all the above-mentioned ingredients, agents become capable of dialogue, exploiting a type of rational and symbolic intelligence that allows arguing and reasoning on the acquired knowledge. They become aware of the context – normative and social – in which they are immersed and through the argumentative process and the dialogue they can decide the next actions be implemented. The normative context is properly represented by the e-institution abstraction. Distributed (micro-)intelligence embodies the enabler of conversation between entities in the MAS infrastructure, making feasible the computation of the law and its incarnation in a computational system.
Dialogue is not expected only between agents but also with the environment and in particular
with the e-institution. The bi-directional dialogue and interaction process, as well as enabling
negotiation between system entities, also allow a prompt examination of each agent and of
their status in order to detect automatically and real-time a possible deviation from the rules
or policies envisaged. The issue of reaching a consensus in an ensemble of cooperating and
interacting autonomous components by distributed negotiations has been already investigated
in the field of multi-agent systems [
Figure 2 (left) summarises our vision by highlighting the main roles involved in the system as well as the two main activity flows—from data to users (left) and the opposite, from user goals and desires to activity planning, and lower-level commands (right). The grey boxes, that we describe more in detail in the following section, represent the technologies involved in the vision, while arrows represent the expected provided functionalities.
On one side, knowledge is collected from various sources – e.g., domain-specific knowledge, ontologies, sensors raw data – and is then exploited by agents that live in a normative environment (e-institutions). Agents’ activity is then regulated by norms but also addicted to situated knowledge over which agents can reason and discuss in order to achieve goals. The multi-agent system, also thanks to its rational reasoning and argumentation capabilities, can provide outcomes to the users as well as explanations for their behaviours. On the other side, humans can insert input into the system – like desires, preferences, or goal to achieve – and these are transposed into agents’ goal, corresponding activity planning, and lower-level commands for actuators.
The law, in this vision, become an internal component of the computational processes: legal norms, values, and principles are mapped and translated into, a computable representation of legal information – i.e., into the system knowledge – that is directly processed by computational entities. Agents become artificial legal agents able to comply with legal requirements and to reason over them.
Computable law assumes a vision of compliance/law by design in this architecture combining top-down compliance with predefined rules and bottom-up learning from cases and experience with capability to address regulatory conflicts according to legal values and principles.
The above considerations translate into diferent engineering approaches and therefore diferent nature of algorithms and techniques that can and must be considered when building the system. In the following, we encapsulate the main technologies involved as well as research challenges and opportunities.
Knowledge representation and related techniques are the cornerstone of the envisioned distributed system, to be able to argue and reason over it.
Of course, system knowledge has to take into account domain-specific knowledge and largescale ontologies as repositories to interpret the knowledge bases available to the agents and to reason and argument over it. Knowledge could be continuously modified, adapted, and refined by the agents, according to their experience and perception of the environment or to learning from experience.
Accordingly, the knowledge base is plausible that is assembled by two main sources: on the one hand, ontologies and hand-crafted rules, on the other hand, rules learned from big data. Advances in machine learning will allow extracting knowledge from this data and to merge it with the former. Hybrid approach dealing with the integration of symbolic and sub-symbolic approaches becomes of paramount importance.
In this context, there are several issues and challenges to be tackled, to cite few, automatic extraction of knowledge from ML models, extraction of commonsense knowledge from the context, integration of the diverse knowledge in an appropriate logical language that allows argumentation and inference process to be performed. Several research fields are already facing these issues, but the general problem is far from being solved. For sure, we believe that a suitable integration of symbolic and sub-symbolic approaches can help in the achievement of the construction of proper system knowledge.
In our vision, a fundamental role is played by machine learning involved in diferent phases— namely, data processing & rule learning, and planning.
Data processing & rule learning. At the most straightforward level, machine learning
techniques are clearly involved in raw input data elaboration, coming from sensors and/or documents,
into more complex, high-level, structured information. Moreover, agents should be able to learn
policies from past experience, by adapting both to the changing environment, and to the
continuous progress of the society. Data aggregation, feature extraction, clustering, classification,
data and pattern mining techniques are typically employed today to reach these objectives.
We believe that hybrid approach could provide promising solutions to these tasks, by merging
logic with probabilistic models and statistical learning, so to eficiently handle advantages of
both symbolic and sub-symbolic approaches and moving towards explainable systems [
Planning. Distributed problem solving, planning, reinforcement learning, and cooperation
[
From a more implementation-oriented perspective, given that conversations are a new means of orchestrating the activities of distributed agents, an open research question – and a key one, too – is to understand which services should a middleware provide in order to support such distributed conversations.
The multi-agent infrastructure need not only to allow coordination among system actors
but also include the possibility of customisable and reactive artefacts, capable of incorporating
regulation and norms and micro-intelligence (possibly in the form of service). Moreover, the
middleware should provide support for discussions via an open and shared discussion space,
enabling dialogue among components that do not necessarily know each other in advance, and
also providing services and or techniques for sharing knowledge, e.g., a tuple space [
Argumentation is a necessary feature for agents to talk and discuss to reach an agreement.
Several existing works establish the maturity of argumentation models as a key enabler of our
vision [
Despite the long history of research in argumentation and the many fundamental results
achieved, much efort is still needed to efectively exploit argumentation in our envisioned
framework. First, research on argumentation has mostly been theoretical, practical applications
to real-world scenarios have only recently gained attention and are not yet reified in a
readyto-use technology [
In any case, we think that the concept of locality is crucial and should be considered along with distributed argumentation—coherently with the notion of micro-intelligence. In addition, it could be interesting to envision a framework where the specific argumentation and inference process can be swapped at runtime, making consideration on the specific situation, and introducing, for instance, abductive reasoning, or probabilistic argumentation or the algorithm deemed best and necessary for such a situation. The dialogue, therefore, becomes possible by following diferent rationales that can guide the decision-making process, possibly comparing diferent perspectives and visions.
Finally, techniques coming from natural language processing, computer vision speech recognition – already a reality in most everyday applications (e.g., Amazon Alexa, Google Home,. . . ) – become essential components of our vision for humans interaction. The challenge in the envisioned scenario is always related to the distributed issues, i.e., making commands possibly understandable to a multitude of agents and vice-versa. Existing algorithms should, therefore, be adapted for dealing with distributed and pervasive environments.
Validation through simulation is of paramount importance when dealing with MAS, and even
more when dealing on the integration of many diferent technologies—as in our vision.
Simulation is concerned with the truthfulness of a model with respect to its problem domain and the
correctness of its construction. In other words, simulation calls for verification, i.e., “building
the system right" and validation, i.e., “building the right system" [
In the trafic management scenario, for instance, by simulating car-following in continuous
trafic flow and comparing simulation data with the data collected from the actual road, the
reliability of the model and the architecture. Many simulation models have been proposed on
the topic [
Also in this field, the simulation models must be able to mix ingredients from the techniques above discussed (symbolic and sub-symbolic) and should be adapted to meet the software engineering requirements these techniques refer to.
The heterogeneous nature of intelligence required by pervasive AI systems along with fears related to sub-symbolic techniques more and more exploited in such systems require for models and technologies guaranteeing explainability as one of the main requirement.
Within this context, we believe that argumentation can play a major role, as it allows debates to be studied and analysed, reasoning and persuasion to be exploited in dialogues, with a well-grounded theoretical framework.
In this paper, we show how diferent dialogues can occur in such pervasive contexts, highlighting the advantages of employing argumentation. Interpretability of decision making, tolerance to uncertainty, adaptiveness, robustness of the system, and improved trust by end-users and amongst interacting components, are the most notable benefits of the proposed approach.
The main limit of the argumentation approach is the typical assumption to have an external judge or authority that has to control the whole argumentation process, but this is very unlikely in a dynamic, distributed scenario like the one we propose. This aspect will certainly be the subject of future work.
However, model and techniques that should play a key role in the engineering of intelligent systems – even if with enhancements and extensions – have been discussed and constitute a starting point for further researches.
Roberta Calegari and Giovanni Sartor have been supported by the H2020 ERC Project “CompuLaw” (G.A. 833647). Andrea Omicini has been supported by the H2020 Project “AI4EU” (G.A. 825619).