Skip to main content

Exploring the Unknown through Reincarnation and Co-evolution

Periodic Reporting for period 3 - Phoenix (Exploring the Unknown through Reincarnation and Co-evolution)

Reporting period: 2018-04-01 to 2019-09-30

"What is the issue being addressed?
There is no way to efficiently explore unknown environments. Phoenix addresses this with a method that uniquely combines innovations in hardware, sensing, and artificial evolution to produce swarms of evolving motes that efficiently explore, map and sense in environments which would otherwise remain unknown. We show the success of our project through a demonstration of how Phoenix maps a pipeline.

Why is it important for society?
Whenever humankind explores a new environment, we always acquire new knowledge. That knowledge is intrinsically important because it broadens our collective understanding of natural phenomena and may entail social implications regarding quality of life, health, and wellbeing. Consider the following challenges:

– Mapping pipelines to find obstructions, leaks or faults to more efficiently deliver drinking water, prevent contamination, or monitor water quality deep within pipeline networks.
– Exploring underground channels, which cannot be otherwise accessed without damaging them to more efficiently extract oil or natural gas or to the search of natural CO2 storage locations.
– Measuring from the depths of glaciers or inside volcanos to better model climate change.

However, there are places which, even today, cannot be reached by even the most advanced sensors. Phoenix targets those places.

What is the overall objective?
The objective of Phoenix is to develop a method to explore unknown and inaccessible environments. This involves: 1) the design of versatile agents (i.e. mote) technology, 2) the definition of techniques to formalize expert knowledge to influence the evolution of agents and their ""rebirth"" and 3) the development of a co-evolutionary framework to jointly optimize sensor motes and environment models.
Phoenix also sheds light on emergent properties of self-organization, local adaptation and division of labor in autonomous systems.

Conclusion
Although not all system components are fully integrated and some steps are still manual, the feasibility of Phoenix was confirmed through demonstrators. In the main demonstrator, agents explore an unknown pipe loop system and a human interface provides input to the system, a co-evolution method optimizes the agents based on extracted information from the environment, and uses this to configure instincts in agents. While the hardware in the agents uses off-the-shelf components, the implementation of the instincts, as well as the adaptable parameters of the sensors are made consistent with the developed miniaturized hardware. This enables, in a future step, to seamlessly replace the off-the-shelf components with the miniaturized IC based hardware. Three additional demonstrators have shown the hardware development contributions of Phoenix, aiming for miniaturization and energy reduction. Ultrasound electronics and transducers were implemented that are smaller and more efficient than prior-art, and we validated that communication and ranging can be performed underwater up to several meters of distance. Moreover, adaptable sensors and instincts were integrated and also confirmed their functionality, at a size and energy budget that is orders of magnitude below off-the-shelf components. With these components, it is shown feasible to reduce the size of the black ball agent (6 cm in diameter) to a cm-sized (or smaller) ball, AND to operate the device for a much longer amount of time.
"
During the project, challenges associated with the development of the motes were addressed. The motes explore the unknown environment and record a variety of sensor data during this exploration, for later retrieval. However, due to mote space limitations, energy and storage are very scarce resources. To enable efficient behavior over a wide range of dynamic environments, versatile sensing and compression circuits were implemented. The end result is a highly optimized cm-scale system constructed out of commercial ICs, and a miniaturized integrated system using custom chips. Both implementations are controlled by an online smart adaptation block that executes offline learned behavioral rules for self-adaptation, also called “instincts”. The “instinct” can adapt the mote’s behavior autonomously, dependent on the sensor inputs, to optimize information acquisition versus energy consumption. Moreover, a single integrated sensor interface, ultrasound communication and localization methods were developed for the mote, and prototypes of transducer elements and transceiver electronics were designed and tested.

The use of such versatile physical motes involves expertise. Artificial evolution ensures that the motes’ self-adaptation scheme evolves towards a more efficient exploration of the environment. This requires “high level” reasoning. Accordingly, we introduced an open, extendible knowledge representation scheme to represent and reason on exploration procedures. We developed a knowledge base to store the knowledge needed in these exploration tasks and provided methods to extend our scheme to represent uncertain procedural knowledge and implemented a Human Interface Layer.

Additionally, we developed an Evolutionary Algorithm (EA) which uniquely involves recourse to human knowledge provided to the Phoenix system by experts and users. The result, a Knowledge-Integrated Evolutionary Algorithm (KIEA), is able to identify the intrinsic features of an unknown optimization problem and adjust its parameters and operators to obtain superior optimization performance. KIEA was initially tested on numerical benchmarks and simple simulation scenarios. Moreover, we introduced a novel algorithm for evolving Behavior Trees, which we call an Instinct Evolution Scheme (IES), that includes a solution space analysis step and a procedure to reduce the search space of the agents’ internal behavior. Finally, we integrated the KIEA and IES into a complete co-evolutionary framework where motes and environment models jointly evolve, to collect interesting solutions by using different solution concepts. This may serve as a new tool for evolving the behavior of any agent, from robots to wireless sensors, to address different needs from industry.

In the H2020 Launchpad project SMARBLE the business case for the Phoenix project results is further explored.
On the agent side, state-of-the-art research focused on maximizing data compression with compute-intensive algorithms on one hand, and on lightweight compression schemes which minimize energy spent in the compression operation on the other. The (self-)adaptivity is online controlled using a novel concept of “instincts”. When running typical behavioral trees, the custom processor consumes less than 100nW, while it takes less than 1mm2 in chip area.

The versatile sensor interface, with a proportional scaling of power over 3 to 4 orders of magnitude, achieves state-of-the-art efficiency for each sensing mode and a lowest power consumption of 0.34nW.

The developed ultrasound electronics achieve the best efficiency per communicated bit compared to prior art, and the first FMCW receiver for ultrasound ranging was proposed. Further, the developed PMUT-based transducers prove the ability for underwater communication at low voltage levels (compatible to low-voltage IC integration).
Phoenix pipe-loop test setup
Phoenix diagram
Phoenix motes