Abstract The ‘out-of-place’ update nature of Solid State Drives (SSDs) introduces a security risk. Scrubbing, a secure deletion method, mitigates this issue but negatively affects SSD endurance and requires redundant data for recovery due to page errors. Previously, the RAID-5 based scheme was used to manage redundant data. While effective for reading, it causes significant write latency due to the channel blocking issue. In response, we propose FlashFox, which integrates secret sharing and Reed-Solomon coding into SSDs, enabling the application of scrubbing within an encrypted storage environment. This innovation ensures secure deletion by cleaning up sensitive data, thereby reducing wear on SSD endurance. Furthermore, we have developed a RAID-4-based scheme for implementing FlashFox on SSDs. This scheme, by assigning specific channels to manage redundant data, successfully avoids the channel blocking issue prevalent in RAID-5. Experimental results show FlashFox reduces endurance wear by at least 15% compared to traditional scrubbing methods and writing response delay by at least 8$\times $ compared to RAID-5 based scheme.

The massive computational and memory requirements of deep convolutional neural networks (DCNNs) have led to the development of neural network (NN) accelerators. However, as DCNN models grow in size, the demands on NN accelerators in terms of performance, memory bandwidth, and power efficiency continue to increase. We, therefore, present 9Ring , a flexible and efficient DCNN accelerator that takes full advantage of 3D-stacked memory, focusing on its hardware architecture, software scheduling, and optimization strategy. In particular, we first show that the mismatch between DCNN accelerators and DCNN models can lead to increased energy consumption and performance bottlenecks. We then present three flexible dataflow scheduling strategies to mitigate this mismatch. Afterward, we introduce an energy efficiency analysis tool that can automatically search for the optimal scheduling scheme with respect to different DCNN models for energy efficiency. Finally, we conduct an empirical study showing that 9Ring can reduce energy consumption by 31.4% and 43.9% on average, and improve performance by 12% and 10% on average, compared with Tetris and the NN accelerators on conventional low-power DRAM memory systems, respectively.

Over the past four decades, Learning Classifier Systems (LCSs) have faced challenges in producing accurate and interpretable models for domains with continuous features, mainly due to the irrelevance issue caused by genotypic methods. These methods directly modify genotypes (conditions), leading to the creation of irrelevant rules. Phenotypic LCSs, which first modify a rule's phenotype (covered instance set) before altering its genotype, can avoid this issue. However, previous phenotypic LCSs struggle with overfitting, resulting in lower testing performance. In response, we propose a novel phenotypic LCS featuring innovations: 1) a heterogeneous phenotype approach in the rule discovery mechanism to alleviate overfitting, and 2) Informed Mutation leverages the inherent neighbouring of similar instances to enhance rule generalization, thereby improving model interpretability. The proposed LCS demonstrates its success with superior testing performance and more interpretable models in all experiments compared to other LCSs. Notably, in a problem with 2048 features, the proposed LCS model outperformed the genotypic UCS by achieving a 97.4% testing accuracy with just 13 rules, compared to the UCS's 9961 rules but only 49.9% accuracy.

Learning Classifier Systems (LCSs) are a paradigm of rule-based evolutionary computation (EC). LCSs excel in data-mining tasks regarding helping humans to understand the explored problem, often through visualizing the discovered patterns linking features to classes. Due to the stochastic nature of EC, LCSs unavoidably produce and keep redundant rules, which obscure the patterns. Thus, rule compaction methods are invoked to produce a better population by removing problematic rules. Previously, compaction methods have neither been tested on large-scale problems nor been assessed on the performance of capturing patterns. We review and test the most popular compaction algorithms, finding that across multiple LCSs’ populations for the same task, although the redundant rules can be different, the accurate rules are common. Furthermore, the patterns contained consistently refer to the nature of the explored domain, e.g., the data distribution or the importance of features for determining actions. This extends the [ O ] set hypothesis proposed by Butz et al. [1], in which an LCS is expected to evolve a minimal number of non-overlapped rules to represent an addressed domain. Two new compaction algorithms are introduced to search at the rule level and the population level by compacting multiple LCSs’ populations. Two visualization methods are employed for verifying the interpretability of these populations. Successful compaction is demonstrated on complex and real problems with clean datasets, e.g., the 11-bits Majority-On problem that requires 924 different interacting rules in the optimal solution to be uniquely identified to enable correct visualization. For the first time, the patterns contained in learned models for the large-scale 70-bits Multiplexer problem are visualized successfully.

Learning Classifier Systems (LCSs) are a group of rule-based evolutionary computation techniques, which have been frequently applied to data-mining tasks. Evidence shows that LCSs can produce models containing human-discernible patterns. But, traditional LCSs cannot efficiently discover consistent, general rules - especially in domains that have unbalanced class distribution. The reason is that traditional search methods, e.g. crossover, mutation, and roulette wheel deletion, rely on stochasticity to find and keep optimum rules. Recently, absumption has been introduced to deterministically remove over-general rules, which is complementary to subsumption that deterministically removes over-specific rules. It is hypothesized that utilizing just assumption & subsumption transforms the search process from stochastic to deterministic, which benefits LCSs in evolving interpretable models and removing the need to tune search parameters to the problem. Interrogatable artificial Boolean domains with varying numbers of attributes are considered as benchmarks. The new LCS, termed Absumption Subsumption Classifier System (ASCS), successfully produces interpretable models for all the complex domains tested, whereas the non-optimal rules in existing techniques obscure the patterns. ACSC's ability to handle complex search spaces is observed, e.g. for the 14-bits Majority-On problem the required 6435 different cooperating rules were discovered enabling correct pattern visualization.

Learning Classifier Systems (LCSs), a 40-year-old technique, evolve interrogatable production rules. XCSs are the most popular reinforcement learning based LCSs. It is well established that the subsumption method in XCSs removes overly detailed rules. However, the technique still suffers from overly general rules that reduce accuracy and clarity in the discovered patterns. This adverse impact is especially true for domains that are containing accurate solutions that overlap, i.e. one data instance is covered by two plausible, but competing rules. A novel method, termed absumption, is introduced to counter over-general rules. Complex Boolean problems that contain epistasis, heterogeneity and overlap are used to test the absumption method. Results show that absumption successfully improves the training performance of XCSs by counteracting over-general rules. Moreover, absumption enables the rule-set to be compacted, such that underlying patterns can be precisely visualized successfully. Additionally, the equations for the optimal size of solutions for a problem domain can now be determined.

It has been shown that identifying building blocks of knowledge and then reusing them to solve complex problems is a practical and useful endeavor. Previous work made it possible to solve various, until then, intractable tasks. However, the individual algorithms targeted one specific problem type, e.g. scalable problems or domains with repeating patterns. The question that arises is: Can the disparate techniques be combined into a single approach to solve more complex problems that span several domains or that may be unknown to the agent? The first stage in developing such a system is to be able to recognise domains from unidentified input stimuli and identify the approaches best suited to them. The novel work here aims to realise this primary stage by combining several code-fragment (CF) based XCS systems. The stimulus and its guiding effect, will be instrumental in helping the agent decide which of its stored systems is the most capable of solving the problem, or if there is a conflict between possible solutions. Importantly, the agent will be capable of determining if the current problem is entirely new, in which case it spawns a training agent to produce a tractable solution to store and reuse. The proposed technique relies on the proven benefits in scalability of CF based systems and furthers the body of knowledge by tackling unknown problems (to the agent). The main contribution of this research is that a system of proven CF techniques is used for the first time. We show that by utilizing the new CF system, it is possible to identify an unknown problem and to arrive at a viable solution.