IEEE Technology and Society Magazine - June 2020 - 54

through meta-deliberation, i.e., considering the effectiveness of the deliberative properties of the system;
and by requiring periodic deliberative scrutiny of any
stable institutional arrangements.
This work proposes that these processes can be
specified as algorithms.

The idea behind algorithmic reflexive governance is to
specify and implement algorithms for each of the deliberative processes defining the four dimensions of reflexivity. This specification of deliberative processes as
algorithms leverages several prior works from self-organizing systems, including: studies of relevant expertise
aggregation [5] to join expert and lay decision-making;
interactional justice [6] as a way of achieving meta-consensus; situating a system-of--systems within the zone of
dignity [7] as a form of meta-deliberation, and the use of
quasi-stability [8] to scrutinize and (if necessary) restore
stability. We examine each of these processes in turn.

A formal algorithmic specification, and its implementation using multi-agent systems, was used to select policies for regulating community energy systems [9] (e.g.,
given uncertainty in renewable energy generation, how
much energy to import, export, generate, store, etc.). In
this simulation, there are a set of criteria for evaluating
policies, and for each criterion there is a set of metrics
evaluating it. Each agent "knows" a subset of criteria,
and a subset of the related metrics. At regular intervals,
a number of "expert" groups, comprising a subset of the
agents, were formed. Given a starting policy (which
could either be an initial policy or some random policy),
each group of agents used machine learning algorithms
to produce a recommendation for the "best" policy
option under the existing and expected operational conditions. There were two approaches to policy optimization: off-line evolution, which used genetic programming
(GP) to discover an optimal policy a priori; and on-line
evolution, which used GP and reinforcement learning to
select different operating policies at run-time.

Relevant Expertise Aggregation (REA)

Interactional Justice

In the first dimension, relevant expertise aggregation
(REA) is proposed in order to address the tension
between majority preference and expert judgement [5].
REA is a process of decision-making that is both "democratic" (in the sense that it takes into account the
expressed preferences of the majority) and "epistemic" (in
the sense that it takes into account the knowledge and
judgement of experts). The deliberative process is that
expert groups make recommendations and citizen groups
select policy priorities based on those recommendations.
A procedure is informally described in [5]. A governing institution is faced with a set of problems that
need to be addressed. For each problem, it identifies a
limited number of domains: each domain is ranked in
[0,1] and weighted (with the sum of weights on the
domains normalized to 1) according to its perceived
relevance to the issue. A group of experts in each
domain are consulted: it is supposed that these
experts are reliable, i.e., they have good reasons to
disclose their true opinion, and have a probability better than random of reaching the "correct" decision
(taking advantage of the Condorcet Jury Theorem
[10]). The group of domain experts is offered a limited
number of policy options, each option being understood as a potential solution to the issue. Each option
is given a final score as a proportion of the expert
votes multiplied by the weight of the domain, and
summed over all domains. The final scores are used
by the governing/decision making body to determine
policy selection. This can be done a number of different ways, depending on political regime (see Sidebar:
"Relevant Expertise Aggregation").

For the second dimension of reflexivity, interactional justice might address the tension between diversity and consensus. Interactional justice in sociology [11] and
organizational theory [12] deals with the interpersonal
treatment of the members of an institution according to
the outcomes produced by its rules, structures, and decision-makers. Therefore, interactional justice is a user-centric aspect of justice that can be used for realizing values
of fairness and inclusivity in organizations and communities. It has been shown how the concept of interactional
justice can be applied to self-organizing electronic institutions composed of computational components (i.e., software agents), where it is aimed at realizing shared values
amongst a set of disparate agents who are members of
the same (electronic) institution [6].
This use of interactional justice in self-organizing electronic institutions was composed of three phases of interaction in the context of institutional rules. In the first
phase, a subjective evaluation was made of the interpersonal treatment of the individual vis-a-vis its values and
the outcomes of the institution's decisions, i.e., each agent
computed a subjective assessment of the "quality" of an
institution with respect to personal values. In the second
phase, these individual assessments were pooled to form
a collective assessment of the institution's quality with
respect to the collective's shared values (and the priorities
on them); a social network enabled communication
between individuals to disseminate and compare personal treatments by the institution with respect to these
shared values. The third phase addresses how this group
of agents react if they collectively think that their treatment has been "unfair": in particular, reformation of the

Algorithmic Reflexive Governance

IEEE TECHNOLOGY AND SOCIETY MAGAZINE

∕

JUNE 2020

IEEE Technology and Society Magazine - June 2020

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - June 2020