Predictive prosthetic socket design: part 2-generating person-specific candidate designs using multi-objective genetic algorithms.

In post-amputation rehabilitation, a common goal is to return to ambulation using a prosthetic limb, suspended by a customised socket. Prosthetic socket design aims to optimise load transfer between the residual limb and mechanical limb, by customisation to the user. This is a time-consuming process, and with the increase in people requiring these prosthetics, it is vital that these personalised devices can be produced rapidly while maintaining excellent fit, to maximise function and comfort. Prosthetic sockets are designed by capturing the residual limb's shape and applying a series of geometrical modifications, called rectifications. Expert knowledge is required to achieve a comfortable fit in this iterative process. A variety of rectifications can be made, grouped into established strategies [e.g. in transtibial sockets: patellar tendon bearing (PTB) and total surface bearing (TSB)], creating a complex design space. To date, adoption of advanced engineering solutions to support fitting has been limited. One method is numerical optimisation, which allows the designer a number of likely candidate solutions to start the design process. Numerical optimisation is commonly used in many industries but not prevalent in the design of prosthetic sockets. This paper therefore presents candidate shape optimisation methods which might benefit the prosthetist and the limb user, by blending the state of the art from prosthetic mechanical design, surrogate modelling and evolutionary computation. The result of the analysis is a series of prosthetic socket designs that preferentially load and unload the pressure tolerant and intolerant regions of the residual limb. This spectrum is bounded by the general forms of the PTB and TSB designs, with a series of variations in between that represent a compromise between these accepted approaches. This results in a difference in pressure of up to 31 kPa over the fibula head and 14 kPa over the residuum tip. The presented methods would allow a trained prosthetist to rapidly assess these likely candidates and then to make final detailed modifications and fine-tuning. Importantly, insights gained about the design should be seen as a compliment, not a replacement, for the prosthetist's skill and experience. We propose instead that this method might reduce the time spent on the early stages of socket design and allow prosthetists to focus on the most skilled and creative tasks of fine-tuning the design, in face-to-face consultation with their client.

Learning to learn with active adaptive perception.

Increasingly, autonomous agents will be required to operate on long-term missions. This will create a demand for general intelligence because feedback from a human operator may be sparse and delayed, and because not all behaviours can be prescribed. Deep neural networks and reinforcement learning methods can be applied in such environments but their fixed updating routines imply an inductive bias in learning spatio-temporal patterns, meaning some environments will be unsolvable. To address this problem, this paper proposes active adaptive perception, the ability of an architecture to learn when and how to modify and selectively utilise its perception module. To achieve this, a generic architecture based on a self-modifying policy (SMP) is proposed, and implemented using Incremental Self-improvement with the Success Story Algorithm. The architecture contrasts to deep reinforcement learning systems which follow fixed training strategies and earlier SMP studies which for perception relied either entirely on the working memory or on untrainable active perception instructions. One computationally cheap and one more expensive implementation are presented and compared to DRQN, an off-policy deep reinforcement learner using experience replay and Incremental Self-improvement, an SMP, on various non-episodic partially observable mazes. The results show that the simple instruction set leads to emergent strategies to avoid detracting corridors and rooms, and that the expensive implementation allows selectively ignoring perception where it is inaccurate.

Re-inspiring the genetic algorithm with multi-level selection theory: multi-level selection genetic algorithm.

Genetic algorithms are integral to a range of applications. They utilise Darwin's theory of evolution to find optimal solutions in large complex spaces such as engineering, to visualise the design space, artificial intelligence, for pattern classification, and financial modelling, improving predictions. Since the original genetic algorithm was developed, new theories have been proposed which are believed to be integral to the evolution of biological systems. However, genetic algorithm development has focused on mathematical or computational methods as the basis for improvements to the mechanisms, moving it away from its original evolutionary inspiration. There is a possibility that the new evolutionary mechanisms are vital to explain how biological systems developed but they are not being incorporated into the genetic algorithm; it is proposed that their inclusion may provide improved performance or interesting feedback to evolutionary theory. Multi-level selection is one example of an evolutionary theory that has not been successfully implemented into the genetic algorithm and these mechanisms are explored in this paper. The resulting multi-level selection genetic algorithm (MLSGA) is unique in that it has different reproduction mechanisms at each level and splits the fitness function between these mechanisms. There are two variants of this theory and these are compared with each other alongside a unified approach. This paper documents the behaviour of the two variants, which show a difference in behaviour especially in terms of the diversity of the population found between each generation. The multi-level selection 1 variant moves rapidly towards the optimal front but with a low diversity amongst its children. The multi-level selection 2 variant shows a slightly slower evolution speed but with a greater diversity of children. The unified selection exhibits a mixed behaviour between the original variants. The different performance of these variants can be utilised to provide specific solvers for different problem types when using the MLSGA methodology.