IEEE Computational Intelligence Magazine - May 2022 - 54

... embodied intelligence enables agents to adapt to
diverse forms, shapes and environments, hence setting
the stage for the efficacy of EMT with implicit or
explicit genetic transfer to be naturally realized.
Table I summarizes the RMSE values obtained by MFGP
and its single-task counterpart SL-GEP [72]. Superior
results are highlighted in bold. As can be seen, MFGP outperformed
SL-GEP in all experimental settings. Particularly,
the best results of CO2 and DRP were achieved when
paired with their corresponding simplified problem variants.
This is intuitively agreeable as the simplified tasks (generated
by subsampling) are expected to be similar to the original
problem instances, hence engendering fruitful transfers
of genetic building-blocks that speed up convergence and
improve performance.
B. Category 2: EMT in Evolving Embodied Intelligence
Evolutionary robotics takes a biologically inspired view of the
design of autonomous machines [73]. In particular, EAs are
used to adapt robots/agents to their environment by optimizing
the parameters and architecture of their control policy (i.e.,
the function transforming their sensor signals
to motor commands) while accounting for, or
even jointly evolving, the morphology of the
agent itself. It is the design of intelligent
behavior through this interplay between an
agent and its environment, mediated by the
physical constraints of the agent's body, sensory
and motor system, and brain, that is regarded
as embodied intelligence [74]. Put differently, while mainstream
robotics seeks to generate better behavior for a given agent,
embodied intelligence enables agents to adapt to diverse forms,
shapes and environments, hence setting the stage for the efficacy
of EMT with implicit or explicit genetic transfer to be naturally
realized [75].
Imagine evolving embodied intelligence by means of different
tasks parameterized by an agent's morphological and
environmental descriptors. In [76], a multitasking analogue of
an archive-based exploratory search algorithm [77] was used
to train a 6-legged robot to walk forward as fast as possible
under different morphologies derived by changing the
lengths of its legs. Each set of lengths thus defined a specific
task. The experiments evolved walking gait controllers for
2000 random morphologies (or tasks) at once, under the
intuition that a particular controller might transfer as a good
starting point for several morphologies. The results
Smart Home
FIGURE 1 Cloud computing platforms house black-box optimization services where users can simply upload their raw data to have optimized
predictive models delivered [57]. In this setting, EMT could harness knowledge transfers across non-identical but related tasks (e.g., with different
training data and/or device requirements) to enable efficient model configuration.
54 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2022

IEEE Computational Intelligence Magazine - May 2022

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