AI Is Coming for Mathematics, Too?

Consider this.

AI Is Coming for Mathematics, Too

The New York Times

Siobhan Roberts, July 2, 2023

An artificial intelligence (AI)-driven transformation of mathematics looms, with former Google computer scientist Christian Szegedy forecasting computers will match or surpass human mathematicians' problem-solving ability by 2026. Terence Tao at the University of California, Los Angeles said mathematicians' concerns about AI potentially threatening mathematical aesthetics or their profession have emerged in the last several years. The University of Wisconsin-Madison's Jordan Ellenberg suggested AI gadgets could help optimize mathematicians' work. Microsoft's open source Lean proof assistant, which uses automated reasoning powered by AI, is drawing interest for its recent achievements, yet its frequent complaining of being unable to understand the mathematician's inputs makes research awkward. Geordie Williamson at Australia's University of Sydney said mathematicians and computer scientists should participate in discussions about AI's mathematical implications more aggressively.  ,,, ' 

Mixed Integer Linear Programming (MILP)

 Spent most of my early optimization career doing this kind of problem.    Here a technical introduction. Still has lots of value.   I post it here both for nostalgic and real life problem solving uses. Lots of software out there to help too.  Have you used it? 

Mixed Integer Linear Programming: Formal definition and solution space  in TowardsDataScience.   By István Módos 

Theoretical foundations on how MILP is defined and how its solution space looks like visually

This post is a continuation of the series (previous post) on the theory and applications of Mixed Integer Linear Programming (MILP). Today, we look at  the formal, general definition of MILP,

how its space of feasible solutions looks like.

The first part describes how MILP models look in general, what are and are not valid MILP expressions. The second part shows a little bit of theory behind MILP, which will be useful in the future posts when we will be talking about the actual algorithm for solving MILP  ....

 This is quite a long post, so grab a coffee and dive in!  ... ' 

----------------------------------------------------------

ChatGPT Says:

MILP stands for Mixed-Integer Linear Programming, which is a type of mathematical optimization problem. In MILP, the objective function and the constraints are linear, and some of the variables are integer (i.e., they can only take on integer values). The remaining variables are allowed to take on real values.

MILP is a powerful tool for solving a wide range of problems that involve both discrete and continuous decision variables. Examples of applications of MILP include resource allocation, scheduling, production planning, and portfolio optimization.

Solving a MILP problem involves finding the values of the decision variables that optimize the objective function while satisfying the linear constraints. MILP solvers use a variety of algorithms and techniques, such as branch-and-bound, cutting planes, and heuristics, to efficiently search the solution space and find the optimal solution or a good approximation to it.       END

Simplex Algorithm Still Key

Amazes me, was the first thing I learned in engineering analytics methodology to optimize systems  Originally invented for military supply chains in WWII.   Here a technical overview and good historical piece.   

Why the Simplex Method, at Age 75, is Still the Go-To Algorithm?  By Allyn Jackson, Commissioned by CACM Staff, January 31, 2023

ACM NEWS, January 31, 2023, Daniel A. Spielman

Since its birth more than two decades ago, smoothed analysis has been used to analyze the performance of algorithms other than the simplex method, including interior-point methods for linear programming. It also has guided the design of new algorithms.

In 1947, mathematical scientist George Dantzig invented the simplex method, a powerful and practical means to find solutions to linear programming for optimization problems. Scientists lost no time putting the simplex method to use in a variety of applications across government, industry, science, and engineering.

Half a century later, when Daniel A. Spielman was a Ph.D. student at the Massachusetts Institute of Technology, the simplex method stood at the top of the pantheon of important algorithms, yet research had shown the simplex method had proven pitfalls; it ought not to perform as well as it did.  What was going on?

Spielman solved the mystery in 2001, in joint work with Shang-hua Teng, now University Professor and Seeley G. Mudd Professor of Computer Science and Mathematics in the Viterbi School of Engineering at the University of Southern California. Pioneering a technique called smoothed analysis, their work provided an original and compelling explanation of the power of the simplex method and suggested a new paradigm for gauging the effectiveness of algorithms.

This work was recognized in 2008 by the Gödel Prize, sponsored jointly by ACM's Special Interest Group on Algorithms and Computing Theory (SIGACT) and the European Association for Theoretical Computer Science. Now Spielman, Sterling Professor of Computer Science and a professor of Statistics and Data Science and of Mathematics at Yale University, has been awarded the 2023 Breakthrough Prize in Mathematics, in part for this work in optimization.

Spielman has a great ability "to come up with new approaches," said Lance Fortnow, dean of the College of Computing at the Illinois Institute of Technology. "It wasn't like someone else had invented smoothed analysis and Spielman said, 'Let's try it on linear programming'. This was, 'I want to understand linear programming. What's the right way to do it?'"

Simplex Bests Polynomial-time Competition

Dantzig formulated the concept of linear programming as a way to model optimization problems. The model produces a polyhedron, possibly of very high dimension, where the corners represent solutions. The simplex method provides a highly efficient way of moving along polyhedron edges toward an optimal corner. The algorithm was tailor-made for the computing machines that were just beginning to appear when Dantzig did this work.

In the 1970s, the rise of complexity theory brought new precision to the study of efficiency of algorithms. Generally, an algorithm is considered efficient if it runs in polynomial time, meaning that even for the worst-case inputs to the algorithm, the running time is always bounded by a polynomial function of the input size. This is in contrast with algorithms whose running time can increase exponentially with input size.

Soon a curious fact arose: despite its excellent performance in practice, the simplex method is not running in polynomial time. Examples were found on which simplex ran in exponential time. Eventually, polynomial-time algorithms for linear programming were found, but the simplex method continued to be used — and in many situations, outperformed its polynomial-time competitors.

Why does simplex work so well?

This question was in the air when Fortnow was a Ph.D. student at the Massachusetts Institute of Technology (MIT) in the 1980s, several years before Spielman became a graduate student there.  However, he recalls, few were attempting to resolve it. "The simplex method had already been around for 40 years," said Fortnow. The general attitude was, "Well, it works well in practice."

When Spielman and Teng came up with smoothed analysis, it was a big surprise. "It's a whole different way of looking at worst-case complexity," said Fortnow, "and they could apply it to linear programming. Both of these pieces were very exciting."   ... ' 

Empirical Optimization with Divergent Fixed Point Algorithm – When All Else Fails

Thanks to Vincent Granville for sending along his 'Empirical Optimization with Divergent Fixed Point Algorithm – When All Else Fail'       

Entitled “Empirical Optimization with Divergent Fixed Point Algorithm – When All Else Fails”, the full version in PDF format is accessible in the “Free Books and Articles” section, here. Also discussed in details with Python code in his book “Synthetic Data”, available here.  Context him directly here.

While the technique discussed here is a last resort solution when all else fails, it is actually more powerful than it seems at first glance. First, it also works in standard cases with “nice” functions. However, there are better methods when the function behaves nicely, taking advantage of the differentiability of the function in question, such as the Newton algorithm (itself a fixed-point iteration). It can be generalized to higher dimensions, though I focus on univariate functions here.

Perhaps the attractive features are the fact that it is simple and intuitive, and quickly leads to a solution despite the absence of convergence. However, it is an empirical method and may require working with different parameter sets to actually find a solution. Still, it can be turned into a black-box solution by automatically testing different parameter configurations. In that respect, I compare it to the empirical elbow rule to detect the number of clusters in unsupervised clustering problems. I also turned the elbow rule into a fully automated black-box procedure, with full details offered in the same book.

About the author

Vincent Granville is a pioneering data scientist and machine learning expert, co-founder of Data Science Central (acquired by  TechTarget in 2020), founder of MLTechniques.com, former VC-funded executive, author and patent owner. Vincent’s past corporate experience includes Visa, Wells Fargo, eBay, NBC, Microsoft, and CNET. Vincent is also a former post-doc at Cambridge University, and the National Institute of Statistical Sciences (NISS).  

Vincent published in Journal of Number Theory, Journal of the Royal Statistical Society (Series B), and IEEE Transactions on Pattern Analysis and Machine Intelligence. He is also the author of multiple books, including “Intuitive Machine Learning and Explainable AI”, available here. He lives  in Washington state, and enjoys doing research on spatial stochastic processes, chaotic dynamical systems, experimental math and probabilistic number theory. ... '

Empirical Optimization with Divergent Fixed Point Algorithm – When All Else Fails

 Of Technical Interest, worth signing into:

Empirical Optimization with Divergent Fixed Point Algorithm – When All Else Fails

 DECEMBER 9, 2022 EXPLAINABLE AI FEATURED POSTS MACHINE LEARNING SYNTHETIC DATA VISUALIZATION

Entitled “Empirical Optimization with Divergent Fixed Point Algorithm – When All Else Fails”, the full version in PDF format is accessible in the “Free Books and Articles” section, here. Also discussed in details with Python code in my book “Synthetic Data”, available here.

While the technique discussed here is a last resort solution when all else fails, it is actually more powerful than it seems at first glance. First, it also works in standard cases with “nice” functions. However, there are better methods when the function behaves nicely, taking advantage of the differentiability of the function in question, such as the Newton algorithm (itself a fixed-point iteration). It can be generalized to higher dimensions, though I focus on univariate functions here.

Perhaps the attractive features are the fact that it is simple and intuitive, and quickly leads to a solution despite the absence of convergence. However, it is an empirical method and may require working with different parameter sets to actually find a solution. Still, it can be turned into a black-box solution by automatically testing different parameter configurations. In that respect, I compare it to the empirical elbow rule to detect the number of clusters in unsupervised clustering problems. I also turned the elbow rule into a fully automated black-box procedure, with full details offered in the same book .... '

About the author

Vincent Granville is a pioneering data scientist and machine learning expert, co-founder of Data Science Central (acquired by  TechTarget in 2020), founder of MLTechniques.com, former VC-funded executive, author and patent owner. Vincent’s past corporate experience includes Visa, Wells Fargo, eBay, NBC, Microsoft, and CNET. Vincent is also a former post-doc at Cambridge University, and the National Institute of Statistical Sciences (NISS).  

Vincent published in Journal of Number Theory, Journal of the Royal Statistical Society (Series B), and IEEE Transactions on Pattern Analysis and Machine Intelligence. He is also the author of multiple books, including “Intuitive Machine Learning and Explainable AI”, available here. He lives  in Washington state, and enjoys doing research on spatial stochastic processes, chaotic dynamical systems, experimental math and probabilistic number theory.  

Optimal Decision Making

 Technical, but quite interesting point being made.  Optimality may be a good thing,  but how do I embed it in useful real time decisions?  Notably too the consideration of noise, often a key consideration.  This worth a look. 

Algorithm for Optimal Decision-Making Under Heavy-Tailed Noisy Rewards

Chung-Ang University (South Korea), November 17, 2022

Researchers at South Korea's Chung-Ang University (CAU) and Ulsan Institute of Science and Technology created an algorithm that supports minimum loss under a maximum-loss scenario (minimax optimality) with minimal prior data. The algorithm addresses sub-optimal performance for heavy-tailed rewards by algorithms designed for stochastic multi-armed bandit (MAB) problems. CAU's Kyungjae Lee said the researchers proposed minimax optimal robust upper confidence bound (MR-UCB) and adaptively perturbed exploration (MR-APE) methods. The team obtained gap-dependent and independent upper bounds of the cumulative regret, then assessed their methods via simulations conducted under Pareto and Fréchet noises. The researchers found MR-UCB outperformed other exploration techniques with stronger robustness and a greater number of actions under heavy-tailed noise; MR-UCB and MR-APE also could solve heavy-tailed synthetic and real-world stochastic MAB problems.

Full Article

Learning Material

Mechanical Neural Networks:  Structural system of tunable beams  '    Novel idea. 

ACM TECHNEWS

AI-Powered Material Can Learn Behaviors, Adapt

By Interesting Engineering, October 21, 2022

Mechanical engineers at the University of California, Los Angeles have developed an artificial intelligence-powered material that learns behaviors over time, and can adjust to changing circumstances.

The so-called mechanical neural network (MNN) features a structural system of independently tunable beams arranged in a triangular lattice pattern., The researchers said each beam consists of a "voice coil, strain gauges, and flexures that enable the beam to change its length, adapt to its changing environment in real time, and interact with other beams in the system."

An optimization algorithm uses strain-gauge data to calculate rigidity values to govern the network's adaptation, determining how much force should be applied. Cameras on the MNN's outer nodes check the strain-gauge system's validity.

From Interesting Engineering

View Full Article   

Monte Carlo Simulation

 Good intro piece to a method we used for many purposes in the enterprise.   Even creating usable models for key processes that were used for decades.  Consider its similarities to Digital Twins.  

 Monte Carlo Simulation

Darío Weitz   in Towards Data Science, Engineer, Ms. Sc., Former Associate Professor at Ing. en Sistemas de Información, Fac. Reg. Rosario, Univ. Tecnológica Nacional, Argentina. Data Viz Consultant.

Part 1: The News Vendor Problem

In the first article of this series, we defined simulation as a numerical technique consisting of building a mathematical and/or logical model of the system under study and then experimenting with it, collecting data that allows us to obtain an estimator to help solve a complex decision problem.

In the same article, we defined a model as a simplified but valid representation of a real process or system, intending to gain some understanding of its behavior.

We also made a classification of models, distinguishing in particular between continuous models, those in which their behavior (state variables) changes continuously over time, and discrete models, those in which the state variables only change at separate points in time. Another important classification involves static models, those that are a representation of the system at a particular time, and dynamic models, those that evolve over time.

Related to the above classification there are three different types of simulations: continuous event simulation, discrete event simulation, and Monte Carlo simulation.

Principles and concepts about Discrete Event Simulation (DES) were provided in the previously indicated series. We coded several examples with SimPy, an object-oriented, process-based, discrete-event simulation framework based on pure Python. In future articles, we will develop concepts and principles related to continuous event simulation.

In this article (and probably in a couple of others) we are dealing with Monte Carlo Simulations.

Monte Carlo Methods

Monte Carlo Methods (MCM) is a collection of numerical methods for the solution of mathematical problems, where the use of random samples differentiates them from equivalent methods.

The term was coined by the Greek-American physicist Nicholas Metropolis when he was working at Los Alamos National Laboratory with John von Neumann and Stanislaw Ulam in the development of the first atomic bomb. The term gets its origin from the famous casino located in the Principality of Monaco.

The conceptual idea of the MCM consists in the estimation of certain quantities through repeated sampling from models represented in a computer. Two classes of mathematical problems are usually solved with these techniques: integration and optimization.

Concerning the contents described in this series of articles, when we refer to Monte Carlo Simulation models we are talking about static, discrete, stochastic models trying to solve an optimization problem.

From a methodological point of view, a Monte Carlo simulation is a sampling experiment whose aim is to estimate the distribution of a quantity of interest that depends on one or more stochastic input variables. We are particularly interested in calculating point estimates and confidence intervals for that quantities. Inevitably, our estimator will have a sampling error and our first task will be to determine the number of replications to improve the degree of certainty in the value of the estimator.  .... ' 

A Future for AR

 Like the proposed connection to optimization ...   visualizing the value always useful. 

Intensity Control of Projectors in Parallel—Doorway to an Augmented Reality Future

Tokyo Tech News, March 16, 2022

Scientists at Japan's Tokyo Institute of Technology (Tokyo Tech) have developed a method for interacting with dynamic objects in augmented reality (AR), while reducing latency. Dynamic projection mapping combines cameras and projectors that respectively detect and project onto target surfaces, and the researchers' technique can calculate the intensity of each pixel on a target in parallel. This reduces the need for a single large optimization calculation, significantly boosting mapping speed and shortening latency. "The presented high-speed multi-projection is expected to be a major part of important base technologies that will advance spatial AR to derive more practical uses in our daily life," explained Tokyo Tech's Yoshihiro Watanabe. .... ' 

Optimizing Robots with Evolution

Home/News/A Novel Way to Optimize Robots/Full Text

ACM NEWS

A Novel Way to Optimize Robots

By The Economist   November 3, 2021

It might sound obvious that if you want to improve a robot's software, you should improve its software. Agrim Gupta of Stanford University, however, begs to differ. He thinks you can also improve a robot's software by improving its hardware—that is, by letting the hardware adapt itself to the software's capabilities.

As they describe in Nature Communications, he and his colleagues have devised a way of testing this idea. In doing so, they have brought to robotics the principles of evolution by natural selection. They have also cast the spotlight on an evolutionary idea that dates from the 1890s, but which has hitherto proved hard to demonstrate.

There is a wrinkle. The team's robots, which they dub "unimals", are not things of metal and plastic. Rather, they are software entities that interact with a virtual environment in the way that metal-and-plastic devices might interact with a real one. Unimals are pretty simple, having spheres for heads and cylinders for limbs (see picture). The environments through which they roamed were also simple, and came in three varieties: flat arenas, arenas filled with hills, steps and rubble, and ones that had the complexities of the second sort, but with added props like cubes that needed to be moved around.

From The Economist

View Full Article

Simulation with Python

By far my biggest effort in the enterprise was doing simulations.   Often integrated with optimizations, and later AI applications, both knowledge based and neural approaches.    Nice to see how it can be integrated with Python.   We used IBM's forms,  GPSS  and Simscript; Sometimes specified by clients.  Below an intro, more at the link

Monte Carlo Simulation and Variants with Python

Your Guide to Monte Carlo Simulation and Must Know Statistical Sampling Techniques With Python Implementation

By Tatev Karen

Monte Carlo Simulation is based on repeated random sampling. The underlying concept of Monte Carlo is to use randomness to solve problems that might be deterministic in principle. Monte Carlo simulation is one of the most popular techniques to draw inferences about a population without knowing the true underlying population distribution. This sampling technique becomes handy especially when one doesn’t have the luxury to repeatedly sample from the original population. Applications of Monte Carlo Simulation range from solving problems in theoretical physics to predicting trends in financial investments.

Monte Carlo has 3 main usages: estimate parameters or statistical measures, examine the properties of the estimates, approximate integrals.   ..... ' 

Reinforcement Learning and Entropy

 Have recently been looking at Reinforcement Learning methods.   And this as a form of simulation-optimization for 'Twin' style models that need training.  Berkeley BAIR  makes some points about entropy (disorder) in a recent  article.   (Technical)  Consideringe the application. See the full article, linked to below,  for sufficient detail.

Maximum Entropy RL (Provably) Solves Some Robust RL Problems  By Ben Eysenbach    Mar 10, 2021    Berkeley BAIR  AI

Nearly all real-world applications of reinforcement learning involve some degree of shift between the training environment and the testing environment. However, prior work has observed that even small shifts in the environment cause most RL algorithms to perform markedly worse. As we aim to scale reinforcement learning algorithms and apply them in the real world, it is increasingly important to learn policies that are robust to changes in the environment.

Robust reinforcement learning maximizes reward on an adversarially-chosen environment.

Broadly, prior approaches to handling distribution shift in RL aim to maximize performance in either the average case or the worst case. The first set of approaches, such as domain randomization, train a policy on a distribution of environments, and optimize the average performance of the policy on these environments. While these methods have been successfully applied to a number of areas (e.g., self-driving cars, robot locomotion and manipulation), their success rests critically on the design of the distribution of environments. Moreover, policies that do well on average are not guaranteed to get high reward on every environment. The policy that gets the highest reward on average might get very low reward on a small fraction of environments. The second set of approaches, typically referred to as robust RL, focus on the worst-case scenarios. The aim is to find a policy that gets high reward on every environment within some set. Robust RL can equivalently be viewed as a two-player game between the policy and an environment adversary. The policy tries to get high reward, while the environment adversary tries to tweak the dynamics and reward function of the environment so that the policy gets lower reward. One important property of the robust approach is that, unlike domain randomization, it is invariant to the ratio of easy and hard tasks. Whereas robust RL always evaluates a policy on the most challenging tasks, domain randomization will predict that the policy is better if it is evaluated on a distribution of environments with more easy tasks.

Prior work has suggested a number of algorithms for solving robust RL problems. Generally, these algorithms all follow the same recipe: take an existing RL algorithm and add some additional machinery on top to make it robust. For example, robust value iteration uses Q-learning as the base RL algorithm, and modifies the Bellman update by solving a convex optimization problem in the inner loop of each Bellman backup. Similarly, Pinto ‘17 uses TRPO as the base RL algorithm and periodically updates the environment based on the behavior of the current policy. These prior approaches are often difficult to implement and, even once implemented correctly, they requiring tuning of many additional hyperparameters. Might there be a simpler approach, an approach that does not require additional hyperparameters and additional lines of code to debug? ... " 

Try Solving with Optimization

With all the talk about AI in the air, some of the other fundamental methods are being forgotten.   I spent most of my career using and teaching direct optimization methods in the enterprise.   Below a quick overview.   I am not particularly recommending this particular company, but they did a good job presenting the description.  Consider optimization .... it can be be better, easier to use and more direct than AI for the right applications.  It often uses less data.  It is often used for very complex models.   Every decision problem solver should have it in their capabilities.   As a manager I always ask, have we tried an optimization?   Have said it here many times, here it is again. 

What Must I Do to Use a Solver?

To use a solver, you must build a model of your decision problem that specifies:

The decisions to be made, called decision variables,

The measure to optimize, called the objective,

Any logical restrictions on potential solutions, called constraints.

The solver will find values for the decision variables that satisfy the constraints while optimizing (maximizing or minimizing) the objective. ...  " 

Scale of Autonomous Truck Startup

 Considerable launch, had several people contact me re a recent post agreeing that this will promote the implementation of analytical improvements in supply chains.   Providing more solutions.  Still waiting for these to be generally on the road, and more data on their safety...  

Autonomous truck startup Plus raises $200M to accelerate global commercialization   By Duncan Riley in Silicon Angle   See Plus.ai

Autonomous truck startup Plus announced today it has raised $200 million in new funding to accelerate the global commercialization and deployment of its automated trucking system.

That will include the development of a sales and support network to help fleets integrate its automated trucking system and scale up deployments in the U.S. and China. The Series B round was led by Guotai Junan International, CPE and Wanxiang International Investment, with existing investor Full Truck Alliance also participating.

Founded in 2016 and previously known as Plus.ai, Plus is developing technology to enable large-scale commercialization of autonomous transport. Specializing in full-stack self-driving technology to enable autonomous fleets, Plus works closely with original equipment    ... "

BMW Aims to Manufacture with Quantum

BMW looking at nimbler process control for manufacturing.  Embedded is a prediction for use within the next two years.

BMW Takes First Steps Into Quantum Computing Revolution

CNet, Stephen Shankland, January 27, 2021

BMW is using a Honeywell quantum computer to determine which vehicle components to purchase from which supplier at what time, in order to maintain its production schedules while holding down costs. The quantum computer optimizes the automaker’s choices from numerous options and suboptions. While this marks one of the first real-world uses of quantum computing, BMW's Julius Marcea noted, "Our experts anticipate that it will take some more years until real quantum computers can be used for commercial benefit." Honeywell's Tony Uttley said within the next two years, quantum computers will be able to solve optimization problems classical computers cannot handle. ....

Templates for Computerized Robot Design

Seems a useful approach as we use robot solutions for more applications. Includes both physical designs and control programming.   And Optimization to tasks involved. 

Computer-Aided Creativity in Robot Design

MIT News,  By Daniel Ackerman

Massachusetts Institute of Technology (MIT) researchers have developed a system that simulates and optimizes robot design and control programs. RoboGrammar avoids jumbled structures yielded from arbitrary connections between parts by following rules on component arrangement, or graph grammar, inspired by arthropods. Based on inputs of available components and intended terrain, RoboGrammar defines the problem, drafts possible solutions, and chooses the optimal ones, using graph grammar to design all permutations. Each robot has a Model Predictive Control algorithm to model its movements and select the best design, while a graph heuristic search algorithm iteratively samples and assesses sets of robots, learning which designs execute a specific task better. Said Columbia University’s Hod Lipson, “This work is a crowning achievement in the 25-year quest to automatically design the morphology and control of robots.”

Reinforcement Training is Supervised Learning on Optimized data

 My long time background is in systems optimization.  Quite intriguing claims made that could be very useful.  Ultimately very technical.  .

Reinforcement learning is supervised learning on optimized data  Ben Eysenbach and Aviral Kumar and Abhishek Gupta    Oct 13, 2020

The two most common perspectives on Reinforcement learning (RL) are optimization and dynamic programming. Methods that compute the gradients of the non-differentiable expected reward objective, such as the REINFORCE trick are commonly grouped into the optimization perspective, whereas methods that employ TD-learning or Q-learning are dynamic programming methods. While these methods have shown considerable success in recent years, these methods are still quite challenging to apply to new problems. In contrast deep supervised learning has been extremely successful and we may hence ask: Can we use supervised learning to perform RL?... "

Optimization Assisting the User of Large Databases

Using tools like machine learning to adapt how we do complex operations, like analyzing increasingly large databases for complex uses.  Here work at MIT in this space.  I can see this being used further, to analyze and leverage the context in which the data will be used. 

MIT Is Developing a Tool for Machine Learning-Powered Data Retrieval   Oliver Peckham in DataNami

With the global deluge of data, the opportunities are endless – but so are the challenges. Within five years, the world’s data is estimated to reach 175 zettabytes: enough to fill over 23,000 one-terabyte hard drives for every single person alive. In the context of such a data-driven world, managing and sorting through that data is a task that gets harder by the day, with database and query managers struggling to keep up. Now, researchers from MIT are developing a tool to intelligently assist users of large databases.

“It’s like building a database system for every application from scratch, which is not economically feasible with traditional system designs,” explained MIT Professor Tim Kraska in an interview with MIT’s Adam Conner-Simons. Kraska and his colleagues – from the institute’s Computer Science and Artificial Intelligence Laboratory (CSAIL) – are debuting a design for what they call “instance-optimized systems”: database systems that are able to optimize and reorganize themselves in response to the data types and workloads at hand. 

MIT’s instance-optimized system will be the child of two parents: the “Tsunami” and “Bao” tools. Using machine learning, Tsunami (a successor to “Flood”) interprets user queries to reorganize the layouts of databases. Bao, meanwhile, uses machine learning to intelligently pick the appropriate plan for completing a given query. On their own, Tsunami improved query speed up to tenfold, while Bao-created query plans ran up to 50% faster. When combined: the instance-optimized system.

“Query optimizers have been around for years, but they often make mistakes, and usually they don’t learn from them. That’s where we feel that our system can make key breakthroughs, as it can quickly learn for the given data and workload what query plans to use and which ones to avoid,” Kraska said. “Our hope is that a system like this will enable much faster query times, and that people will be able to answer questions they hadn’t been able to answer before.”   ... "

Ant Algorithm for Commercial Fleets

This kind of bio behavior mimicry was experimented with in some yard applications, found to work better for some path variability.

Ant Algorithms Help Fleet Operators Halve Emissions
The Engineer (U.K.)
July 27, 2020

Researchers at Aston University in the U.K. have developed software that imitates how ants share knowledge, in an effort to help cities and towns reduce emissions and achieve clean air targets. The researchers found that ants can keep a record of the best solutions to problems and update their knowledge similarly to how computer algorithms do so. The researchers were able to improve these ant algorithms to reduce the number of decisions they make and apply that knowledge to city-scale fleet-routing problems. Said Aston's Darren Chitty, "Algorithms based on the foraging behavior of ants have long been used to solve vehicle routing problems, but now we have found how to scale these up to city-size fleets operating over several weeks in much less time than before. It means much larger fleet optimization problems can be tackled within reasonable timescales using software a user can put on their laptop."

Reinforcement Learning for Skill Discovery

Can skills be dsicovered.  That is, a means to find better behavior that leads to prescribed real-world goals?   Here in the Google Research blog,  addressing the use of unsupervised reinforcement learning (RL). Note the determination and inclusion of constraints.   Like in classic optimization problems. Largely technical, but thoughtful positioning.  Considerable links in the article below.

DADS: Unsupervised Reinforcement Learning for Skill Discovery
Friday, May 29, 2020
Posted by Archit Sharma, AI Resident, Google Research

Recent research has demonstrated that supervised reinforcement learning (RL) is capable of going beyond simulation scenarios to synthesize complex behaviors in the real world, such as grasping arbitrary objects or learning agile locomotion. However, the limitations of teaching an agent to perform complex behaviors using well-designed task-specific reward functions are also becoming apparent. Designing reward functions can require significant engineering effort, which becomes untenable for a large number of tasks. For many practical scenarios, designing a reward function can be complicated, for example, requiring additional instrumentation for the environment (e.g., sensors to detect the orientation of doors) or manual-labelling of “goal” states. Considering that the ability to generate complex behaviors is limited by this form of reward-engineering, unsupervised learning presents itself as an interesting direction for RL.

In supervised RL, the extrinsic reward function from the environment guides the agent towards the desired behaviors, reinforcing the actions which bring the desired changes in the environment. With unsupervised RL, the agent uses an intrinsic reward function (such as curiosity to try different things in the environment) to generate its own training signals to acquire a broad set of task-agnostic behaviors. The intrinsic reward functions can bypass the problems of the engineering extrinsic reward functions, while being generic and broadly applicable to several agents and problems without any additional design. While much research has recently focused on different approaches to unsupervised reinforcement learning, it is still a severely under-constrained problem — without the guidance of rewards from the environment, it can be hard to learn behaviors which will be useful. Are there meaningful properties of the agent-environment interaction that can help discover better behaviors (“skills”) for the agents?  ... "   ... '