Santiago Gonzalez

There is now significant historical data available on decision making in organizations, consisting of the decision problem, what decisions were made, and how desirable the outcomes were. Using this data, it is possible to learn a surrogate model, and with that model, evolve a decision strategy that optimizes the outcomes. This paper introduces a general such approach, called Evolutionary Surrogate-Assisted Prescription, or ESP. The surrogate is, for example, a random forest or a neural network… Show more

There is now significant historical data available on decision making in organizations, consisting of the decision problem, what decisions were made, and how desirable the outcomes were. Using this data, it is possible to learn a surrogate model, and with that model, evolve a decision strategy that optimizes the outcomes. This paper introduces a general such approach, called Evolutionary Surrogate-Assisted Prescription, or ESP. The surrogate is, for example, a random forest or a neural network trained with gradient descent, and the strategy is a neural network that is evolved to maximize the predictions of the surrogate model. ESP is further extended in this paper to sequential decision-making tasks, which makes it possible to evaluate the framework in reinforcement learning (RL) benchmarks. Because the majority of evaluations are done on the surrogate, ESP is more sample efficient, has lower variance, and lower regret than standard RL approaches. Surprisingly, its solutions are also better because both the surrogate and the strategy network regularize the decision-making behavior. ESP thus forms a promising foundation to decision optimization in real-world problems. Show less

As the complexity of neural network models has grown, it has become increasingly important to optimize their design automatically through metalearning. Methods for discovering hyperparameters, topologies, and learning rate schedules have lead to significant increases in performance. This paper shows that loss functions can be optimized with metalearning as well, and result in similar improvements. The method, Genetic Loss-function Optimization (GLO), discovers loss functions de novo, and… Show more

As the complexity of neural network models has grown, it has become increasingly important to optimize their design automatically through metalearning. Methods for discovering hyperparameters, topologies, and learning rate schedules have lead to significant increases in performance. This paper shows that loss functions can be optimized with metalearning as well, and result in similar improvements. The method, Genetic Loss-function Optimization (GLO), discovers loss functions de novo, and optimizes them for a target task. Leveraging techniques from genetic programming, GLO builds loss functions hierarchically from a set of operators and leaf nodes. These functions are repeatedly recombined and mutated to find an optimal structure, and then a covariance-matrix adaptation evolutionary strategy (CMA-ES) is used to find optimal coefficients. Networks trained with GLO loss functions are found to outperform the standard cross-entropy loss on standard image classification tasks. Training with these new loss functions requires fewer steps, results in lower test error, and allows for smaller datasets to be used. Loss-function optimization thus provides a new dimension of metalearning, and constitutes an important step towards AutoML. Show less

Metalearning of deep neural network (DNN) architectures and hyperparameters has become an increasingly important area of research. Loss functions are a type of metaknowledge that is crucial to effective training of DNNs, however, their potential role in metalearning has not yet been fully explored. Whereas early work focused on genetic programming (GP) on tree representations, this paper proposes continuous CMA-ES optimization of multivariate Taylor polynomial parameterizations. This approach… Show more

Metalearning of deep neural network (DNN) architectures and hyperparameters has become an increasingly important area of research. Loss functions are a type of metaknowledge that is crucial to effective training of DNNs, however, their potential role in metalearning has not yet been fully explored. Whereas early work focused on genetic programming (GP) on tree representations, this paper proposes continuous CMA-ES optimization of multivariate Taylor polynomial parameterizations. This approach, TaylorGLO, makes it possible to represent and search useful loss functions more effectively. In MNIST and CIFAR-10 benchmark tasks, TaylorGLO finds new loss functions that outperform functions previously discovered through GP, as well as the standard cross-entropy loss, in fewer generations. These functions serve to regularize the learning task by discouraging overfitting to the labels, which is particularly useful in tasks where limited training data is available. The results thus demonstrate that loss function optimization is a productive new avenue for metalearning. Show less

Long training times have increasingly become a burden for researchers by slowing down the pace of innovation, with some models taking days or weeks to train. In this paper, a new, general technique is presented that aims to speed up the training process by using a thinned-down training dataset. By leveraging autoencoders and the unique properties of embedding spaces, we are able to filter training datasets to only include only the samples that matter the most. Through evaluation on a standard… Show more

Long training times have increasingly become a burden for researchers by slowing down the pace of innovation, with some models taking days or weeks to train. In this paper, a new, general technique is presented that aims to speed up the training process by using a thinned-down training dataset. By leveraging autoencoders and the unique properties of embedding spaces, we are able to filter training datasets to only include only the samples that matter the most. Through evaluation on a standard CIFAR-10 image classification task, this technique is shown to be effective. With this technique, training times can be reduced with a minimal loss in accuracy. Conversely, given a fixed training time budget, the technique was shown to improve accuracy by over 50%. This intelligent dataset sampling technique is a practical tool for achieving better results with large datasets and limited computational budgets. Show less

Modern robots collect a wealth of rich sensor data during their operation. While such data allows interesting analysis and sophisticated algorithms, it is simply infeasible to store all the data that is generated. However, collecting only samples of the data greatly minimizes the usefulness of the data. We present CC-LOG, a new logging system built on top of the widely-used Robot Operating System that uses a combination of classification and compression techniques to reduce storage… Show more

Modern robots collect a wealth of rich sensor data during their operation. While such data allows interesting analysis and sophisticated algorithms, it is simply infeasible to store all the data that is generated. However, collecting only samples of the data greatly minimizes the usefulness of the data. We present CC-LOG, a new logging system built on top of the widely-used Robot Operating System that uses a combination of classification and compression techniques to reduce storage requirements. Experiments using the Building-Wide Intelligence Robot, a mobile autonomous mobile platform capable of operating for long periods of time in human-inhabited environments, showed that our proposed system can reduce storage requirements by more than an order of magnitude. Our results indicate that there is significant unrealized potential in optimizing infrastructure commonly used in robotics applications and research. Show less

As wireless sensor networks become more ubiqui- tous in the world, the need for lightweight, resilient time synchro- nization protocols is apparent. Wireless nodes’ internal clocks are subject to drift over time due to manufacturing imperfections and environmental changes. While various protocols have been introduced that attempt to correct for this drift, they each have their own peculiarities and issues.
This paper presents a new protocol, the Sticking Heartbeat Aperture Resynchronization… Show more

As wireless sensor networks become more ubiqui- tous in the world, the need for lightweight, resilient time synchro- nization protocols is apparent. Wireless nodes’ internal clocks are subject to drift over time due to manufacturing imperfections and environmental changes. While various protocols have been introduced that attempt to correct for this drift, they each have their own peculiarities and issues.
This paper presents a new protocol, the Sticking Heartbeat Aperture Resynchronization Protocol (SHARP), that reduces synchronization error and resolves shortcomings of existing pro- tocols. We have implemented and compared SHARP to two exist- ing (and noteworthy) time synchronization protocols, Reference Broadcast Synchronization (RBS) and Simple Synchronization Protocol (SISP), on Atmel ATMega328p based microcontroller platforms with IEEE 802.15.4 Xbee radio modules. We show that SHARP exhibits a higher level of synchronization than SISP (which in turn exhibited much better performance than RBS), while requiring significantly fewer messages. Show less

As the world moves towards greater ubiquity of wireless sensor networks, the need for lightweight, resilient time synchronization protocols is apparent. Wireless nodes’ internal clocks are subject to drift over time due to manufacturing imperfections and environmental changes. Such drift can be detrimental for many systems, especially for those where accurate data timestamping is required. Time synchronization protocols introduce a means by which two or more nodes can wirelessly synchronize… Show more

As the world moves towards greater ubiquity of wireless sensor networks, the need for lightweight, resilient time synchronization protocols is apparent. Wireless nodes’ internal clocks are subject to drift over time due to manufacturing imperfections and environmental changes. Such drift can be detrimental for many systems, especially for those where accurate data timestamping is required. Time synchronization protocols introduce a means by which two or more nodes can wirelessly synchronize their internal clocks.

We have implemented and compared two existing time synchronization protocols, Ref- erence Broadcast Synchronization (RBS) and Simple Synchronization Protocol (SISP), on Atmel ATMega328p based microcontroller platforms with IEEE 802.15.4 Xbee radio mod- ules. We have found that SISP is able to achieve much higher synchronization performance than RBS.

Our goal, however, is to improve upon current time synchronization protocols even fur- ther. Thus, we have developed a new protocol, the Sticking Heartbeat Aperture Resynchro- nization Protocol (SHARP), that aims to reduce synchronization error and resolve shortcom- ings of existing protocols. We show that SHARP exhibits a higher level of synchronization than SISP, while requiring significantly fewer messages.

Additionally, to assist in developing accurate time synchronization protocols, we have performed message transmission / reception latency measurement experiments on the Xbee radios using a logic analyzer. Latency was found to be consistent across test runs, exhibiting a low standard deviation under 100μs. Such latency was also found to strongly follow a Gaussian distribution. These results will be useful in a future implementation of SHARP. Show less

In this article, the challenges of developing intelligent wireless geosystems and the pros and cons of interdisciplinary research are examined. First, the range of current practices in wireless sensor network research projects is summarized by analyzing recent contributions in the literature. Specifically, for each wireless sensor network project, we compile details regarding the scope (e.g., simulation, testbed, real deployment), the type of mote used (e.g., TinyOS, Android, Arduino), and the… Show more

In this article, the challenges of developing intelligent wireless geosystems and the pros and cons of interdisciplinary research are examined. First, the range of current practices in wireless sensor network research projects is summarized by analyzing recent contributions in the literature. Specifically, for each wireless sensor network project, we compile details regarding the scope (e.g., simulation, testbed, real deployment), the type of mote used (e.g., TinyOS, Android, Arduino), and the type of sensors used (e.g., inertial, biological).

Next, reasons why Arduino based wireless devices are perhaps the future of applied wireless sensor network research are discussed. Arduino-based systems are inexpensive, easy to use, readily available, well documented, 100% open-source (hardware and software), object oriented, extendable (using the "shield" concept), and flexible (e.g., in terms of radio choices). Lastly, the pros and cons of interdisciplinary research are discussed; that is, we highlight the struggle between working on difficult, relevant, inherently multi-disciplinary research problems while maintaining scholarship in individual fields. Show less