IEEE Computational Intelligence Magazine - May 2022 - 56

TABLE II Comparison of success rates (in %) achieved by
MFEA-II and a single-task canonical EA (CEA) on different
double pole balancing problem instances. Results are
obtained from [26].
MFEA-II
TASK ls
T1
T2
T3
CEA
{T1, T2}
{T1, T3}
0.60m 27% 30% 30%
0.65m 0% 27% -
0.70m 0% -
7%
{T2, T3}
-
27%
27%
{T1, T2, T3}
47%
37%
17%
or 0.70 m ().T3
TT TT TT
12 13 23
Four resulting MTO settings are denoted as
123
{, },{, },{, }, and {, ,}.TTT The architecture
of the neural network controller (two-layer with ten hidden
neurons) was kept the same for all tasks, thus providing an
inherently unified parameter space for transfer. It is well-known
that the double pole system becomes increasingly difficult to
control as the length of the shorter pole approaches that of the
long pole. However, by simultaneously tackling multiple tasks
with different levels of difficulty, the controllers evolved for
simpler tasks could transfer to help solve more challenging
problem instances efficiently.
This intuition was borne out by the experimental studies in
[26], results of which are also depicted in Table II. A single-task
canonical EA (CEA) could only achieve a success rate of 27%
on task T1 while failing on the more challenging instances T2
and T3. In contrast, the MFEA-II algorithm, equipped with
exactly the same operators as CEA, achieved better performance
across all tasks by virtue of unlocking inter-task skills
transfer. Not only did the success rate of T1 reach 47% (indicating
that useful information could even transfer from challenging
to simpler tasks), but that of T2 and T3 also reached a
maximum of 37% and 27%, respectively.
Base
3
Destination
2
1
5
4
7
11
10
9
FIGURE 3 An illustration of multi-agent path planning. Red stars
denote waypoints between the base station and the destination that
must be visited by a set of UAVs. The flight paths of different UAVs share
similar, and hence transferrable, segments (such as segments 1-to-2
in path p1 and 4-to-5 in path p2, or segments 7-to-8 in path p3 and
9-to-11 in path p4) due to their similar surroundings (e.g., buildings).
6
8
C. Category 3: EMT in Unmanned
Systems Planning
Evolutionary approaches are being used to optimize individual
behaviors in robot swarms and unmanned vehicle systems.
Consider unmanned aerial vehicles (UAVs) as an example. As
their usage increases, UAV traffic management systems would
be needed to maximize operational efficiency and safety [80],
avoiding catastrophes such as collisions, loss of control, etc. In
such settings, each UAV may be viewed as an individual
agent that perceives its surroundings to solve its corresponding
task (e.g., path planning). The communication of
acquired perceptual and planning information to other
UAVs in related environments could then lead to better and
faster decisions collectively. An illustration is depicted in
Fig. 3 where flight paths of different UAVs share similar
straight or bent segments; these priors can be transferred
and reused (as common solution building-blocks) to support
real-time multi-UAV optimization. Explicit EMT offers a
means to this end.
An early demonstration of this idea was presented in [81],
where two different multi-UAV missions were optimized jointly
via the MFEA. The missions were optically distinct. While
the first involved a pair of UAVs flying through two narrow
openings in a barrier, the second involved four UAVs flying
around a geofence of circular planform. The flight paths in
both missions however possessed a hidden commonality. In all
cases, the optimal magnitude of deviation from the straight line
joining the start and end points of any UAV's path was the
same. The MFEA successfully exploited this commonality to
quickly evolve efficient flight paths.
A similar application was carried out in [82] for the path
planning of mobile agents operating in either the same or different
workspaces. It was confirmed that EMT could indeed
lead to the efficient discovery of workspace navigation trajectories
with effective obstacle avoidance. In [83], a multi-objective
robot path planning problem was considered to find solutions
that optimally balance travel time and safety against uncertain
path dangers. Given three topographic maps with distinct terrains,
but bearing similarity in the distribution of obstacles, an
EMT algorithm transferring evolved path information was
shown to converge to sets of shorter yet safer paths quicker
than its single-task counterpart.
1) Case study in safe multi-UAV path planning [84]
As a real-world example, a study on the multi-objective path
planning of five simulated UAVs in a 10 7 km2
#
region in the
southwest of Singapore is presented. The problem is characterized
by uncertainty, stemming from the sparsity of data available
to model key environmental factors that translate into
operational hazards. The objective is thus to minimize travel
distance while also minimizing the probability of unsafe
events (which could be caused by flying through bad weather,
or by loss of control due to poor communication signal
strength). The latter objective is quantified based on a path-integral
risk metric derived in [80]. The resultant bi-objective
56 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2022
Path p1
Path p3
Path p4
Path p2
IEEE Computational Intelligence Magazine - May 2022

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2022
