Experiments
NASCENCE’s experimental work uses a hybrid approach. The hybrid approach defines two domains; a computer domain and a physical domain. We have named the approach ‘Evolution-in-materio’ (EIM). A PC in the computer domain runs an evolutionary search algorithm to look for suitable electrical stimuli that will configure a material in such way that it can output the answer to some computational task. The material is part of the physical domain. The computer and physical domain are connected by hardware that can translate the digital data representation in the computer to output the desired electrical stimulus and read the electrical response from the material back to the computer.
Overview of hybrid computation with ‘in materio’ configuration performed by evolutionary search. The computer configures the application of physical signals to a material and tests the output. A genotype of configuration instructions is subject to evolutionary mutation. Physical output from the materials acts as computational input to the selection process. The configuration loop stops when the computational inputs have found a solution or no more improvement is required or possible.
Computational Tasks
The central idea of evolution-in-materio is that the application of some physical signals to a material can cause it to alter how it affects an incident signal. The modified signal is picked up and a fitness score is assigned depending on how close the modified signal is to the desired response. This fitness is assigned to the member of the population that supplied the configuration signal. Ideally, the material would be able to be reset before the application of new signal or configuration data. This is important since without the ability to reset the material, it may retain a memory from past configurations. This could to its turn lead to the same configuration having different fitness values depending on the history of interactions with the material.
Different classes of computational tasks involving evolutionary search might be suitable for the novel substrates and evolutionary platform developed under the NASCENCE project. We have chosen tasks that are sufficiently studied so that search performances and/or computational ability of the substrate can be benchmarked against other studies. However, it must be remembered that, when developing a new computational substrate, speed and efficacy are less important than demonstrating proof-of-concept success to perform some level of computation or search optimisation. Due to the “blue sky” nature of the research into novel materials, much of our equipment and experiment setups are bespoke and preliminary. Rather than developping high level refinements of search techniques, we focus on showing the potential for particular materials to conduct computation via EIM.
One of our candidate problems is the Travelling Salesman Problem (TSP). The TSP is interesting as the problem requires no input (it can be solved by simply generating permutations of cities to visit and then using this to calculate the solution’s path length) and its solution space increases factorially with the number of cities. This meant that we could rapidly reach reasonable problem instances in terms the size of solution space to search. But there are several other classes of problems, some that represent low-level computation (such as logic circuits) and some that represent high level computation (such as search or control tasks). The following table shows several processing tasks for evolutionary computation and assesses the suitability of these for evolution-in-materio.
PROBLEM |
INPUT DATA LENGTH |
OUTPUT DATA LENGTH |
STANDARD BENCHMARKS |
SUITABILITY SCORE |
COMMENTS |
CURVE-FITTING (SYMBOLIC REGRESSION) |
Small |
Small |
Yes |
4 |
Can be compared with other approaches |
TONE DISCRIMINATORS |
Small, but time dependent |
Small |
No |
5 |
There is previous work with which any solutions discovered by NASCENCE can be compared. |
SIGNAL FILTERES |
Small |
Small |
No |
5 |
Common task in evolution-in-materio |
TRAVELLING SALESMAN PROBLEM |
Small |
Large |
Yes |
3 |
Some limitations in terms of number of cities that can be tackled. |
BIN PACKING |
Small |
Small |
Yes |
2 |
Similar limitations to TSP. |
MOTOR CONTROL / OBJECT AVOIDANCE |
Medium |
Medium |
No |
4 |
Requires simulated environment to provide feedback to evolved controller. |
POLE BALANCING / INVERTED PENDULUM |
Small |
Small |
Yes |
3 |
Difficult task to represent if memory is required. |
CLASSIFICATION TASKS |
Large |
Medium |
Yes |
3 |
Extensive library of examples to use. |
FUNCTION OPTIMISATION TASKS |
Medium |
Small |
Yes |
3 |
|
PHYSICAL DESIGN VIA SIMULATED ENVIRONMENTS |
Large |
Medium |
No |
2 |
Requires extensive simulated environment. |
ONE BIT / TWO BIT ADDERS OR SIMILAR COMPOSITION OF LOGIC GATES |
Small |
Small |
Yes |
5 |
Widely studied. |