Tag Archives: d

#437564 How We Won the DARPA SubT Challenge: ...

Posted on November 18, 2020 by Android

This is a guest post. The views expressed here are those of the authors and do not necessarily represent positions of IEEE or its organizational units.​

“Do you smell smoke?” It was three days before the qualification deadline for the Virtual Tunnel Circuit of the DARPA Subterranean Challenge Virtual Track, and our team was barrelling through last-minute updates to our robot controllers in a small conference room at the Michigan Tech Research Institute (MTRI) offices in Ann Arbor, Mich. That’s when we noticed the smell. We’d assumed that one of the benefits of entering a virtual disaster competition was that we wouldn’t be exposed to any actual disasters, but equipment in the basement of the building MTRI shares had started to smoke. We evacuated. The fire department showed up. And as soon as we could, the team went back into the building, hunkered down, and tried to make up for the unexpected loss of several critical hours.

Team BARCS joins the SubT Virtual Track
The smoke incident happened more than a year after we first learned of the DARPA Subterranean Challenge. DARPA announced SubT early in 2018, and at that time, we were interested in building internal collaborations on multi-agent autonomy problems, and SubT seemed like the perfect opportunity. Though a few of us had backgrounds in robotics, the majority of our team was new to the field. We knew that submitting a proposal as a largely non-traditional robotics team from an organization not known for research in robotics was a risk. However, the Virtual Track gave us the opportunity to focus on autonomy and multi-agent teaming strategies, areas requiring skill in asynchronous computing and sensor data processing that are strengths of our Institute. The prevalence of open source code, small inexpensive platforms, and customizable sensors has provided the opportunity for experts in fields other than robotics to apply novel approaches to robotics problems. This is precisely what makes the Virtual Track of SubT appealing to us, and since starting SubT, autonomy has developed into a significant research thrust for our Institute. Plus, robots are fun!

After many hours of research, discussion, and collaboration, we submitted our proposal early in 2018. And several months later, we found out that we had won a contract and became a funded team (Team BARCS) in the SubT Virtual Track. Now we needed to actually make our strategy work for the first SubT Tunnel Circuit competition, taking place in August of 2019.

Building a team of virtual robots
A natural approach to robotics competitions like SubT is to start with the question of “what can X-type robot do” and then build a team and strategy around individual capabilities. A particular challenge for the SubT Virtual Track is that we can’t design our own systems; instead, we have to choose from a predefined set of simulated robots and sensors that DARPA provides, based on the real robots used by Systems Track teams. Our approach is to look at what a team of robots can do together, determining experimentally what the best team configuration is for each environment. By the final competition, ideally we will be demonstrating the value of combining platforms across multiple Systems Track teams into a single Virtual Track team. Each of the robot configurations in the competition has an associated cost, and team size is constrained by a total cost. This provides another impetus for limiting dependence on complex sensor packages, though our ranging preference is 3D lidar, which is the most expensive sensor!

Image: Michigan Tech Research Institute

The teams can rely on realistic physics and sensors but they start off with no maps of any kind, so the focus is on developing autonomous exploratory behavior, navigation methods, and object recognition for their simulated robots.

One of the frequent questions we receive about the Virtual Track is if it’s like a video game. While it may look similar on the surface, everything under the hood in a video game is designed to service the game narrative and play experience, not require novel research in AI and autonomy. The purpose of simulations, on the other hand, is to include full physics and sensor models (including noise and errors) to provide a testbed for prototyping and developing solutions to those real-world challenges. We are starting with realistic physics and sensors but no maps of any kind, so the focus is on developing autonomous exploratory behavior, navigation methods, and object recognition for our simulated robots.

Though the simulation is more like real life than a video game, it is not real life. Due to occasional software bugs, there are still non-physical events, like the robots falling through an invisible hole in the world or driving through a rock instead of over it or flipping head over heels when driving over a tiny lip between world tiles. These glitches, while sometimes frustrating, still allow the SubT Virtual platform to be realistic enough to support rapid prototyping of controller modules that will transition straightforwardly onto hardware, closing the loop between simulation and real-world robots.

Full autonomy for DARPA-hard scenarios
The Virtual Track requirement that the robotic agents be fully autonomous, rather than have a human supervisor, is a significant distinction between the Systems and Virtual Tracks of SubT. Our solutions must be hardened against software faults caused by things like missing and bad data since our robots can’t turn to us for help. In order for a team of robots to complete this objective reliably with no human-in-the-loop, all of the internal systems, from perception to navigation to control to actuation to communications, must be able to autonomously identify and manage faults and failures anywhere in the control chain.

The communications limitations in subterranean environments (both real and virtual) mean that we need to keep the amount of information shared between robots low, while making the usability of that information for joint decision-making high. This goal has guided much of our design for autonomous navigation and joint search strategy for our team. For example, instead of sharing the full SLAM map of the environment, our agents only share a simplified graphical representation of the space, along with data about frontiers it has not yet explored, and are able to merge its information with the graphs generated by other agents. The merged graph can then be used for planning and navigation without having full knowledge of the detailed 3D map.

The Virtual Track requires that the robotic agents be fully autonomous. With no human-in-the-loop, all of the internal systems, from perception to navigation to control to actuation to communications, must be able to identify and manage faults and failures anywhere in the control chain.

Since the objective of the SubT program is to advance the state-of-the-art in rapid autonomous exploration and mapping of subterranean environments by robots, our first software design choices focused on the mapping task. The SubT virtual environments are sufficiently rich as to provide interesting problems in building so-called costmaps that accurately separate obstructions that are traversable (like ramps) from legitimately impassible obstructions. An extra complication we discovered in the first course, which took place in mining tunnels, was that the angle of the lowest beam of the lidar was parallel to the down ramps in the tunnel environment, so they could not “see” the ground (or sometimes even obstructions on the ramp) until they got close enough to the lip of the ramp to receive lidar reflections off the bottom of the ramp. In this case, we had to not only change the costmap to convince the robot that there was safe ground to reach over the lip of the ramp, but also had to change the path planner to get the robot to proceed with caution onto the top of the ramp in case there were previously unseen obstructions on the ramp.

In addition to navigation in the costmaps, the robot must be able to generate its own goals to navigate to. This is what produces exploratory behavior when there is no map to start with. SLAM is used to generate a detailed map of the environment explored by a single robot—the space it has probed with its sensors. From the sensor data, we are able to extract information about the interior space of the environment while looking for holes in the data, to determine things like whether the current tunnel continues or ends, or how many tunnels meet at an intersection. Once we have some understanding of the interior space, we can place navigation goals in that space. These goals naturally update as the robot traverses the tunnel, allowing the entire space to be explored.

Sending our robots into the virtual unknown
The solutions for the Virtual Track competitions are tested by DARPA in multiple sequestered runs across many environments for each Circuit in the month prior to the Systems Track competition. We must wait until the joint award ceremony at the conclusion of the Systems Track to find out the results, and we are completely in the dark about placings before the awards are announced. It’s nerve-wracking! The challenges of the worlds used in the Circuit events are also hand-designed, so features of the worlds we use for development could be combined in ways we have not anticipated—it’s always interesting to see what features were prioritized after the event. We test everything in our controllers well enough to feel confident that we at least are submitting something reasonably stable and broadly capable, and once the solution is in, we can’t really do anything other than “let go” and get back to work on the next phase of development. Maybe it’s somewhat like sending your kid to college: “we did our best to prepare you for this world, little bots. Go do good.”

Image: Michigan Tech Research Institute

The first SubT competition was the Tunnel Circuit, featuring a labyrinthine environment that simulated human-engineered tunnels, including hazards such as vertical shafts and rubble.

The first competition was the Tunnel Circuit, in October 2019. This environment models human-engineered tunnels. Two substantial challenges in this environment were vertical shafts and rubble. Our team accrued 21 points over 15 competition runs in five separate tunnel environments for a second place finish, behind Team Coordinated Robotics.

The next phase of the SubT virtual competition was the Urban Circuit. Much of the difference between our Tunnel and Urban Circuit results came down to thorough testing to identify failure modes and implementations of checks and data filtering for fault tolerance. For example, in the SLAM nodes run by a single robot, the coordinates of the most recent sensor data are changed multiple times during processing and integration into the current global 3D map of the “visited” environment stored by that robot. If there is lag in IMU or clock data, the observation may be temporarily registered at a default location that is very far from the actual position. Since most of our decision processes for exploration are downstream from SLAM, this can cause faulty or impossible goals to be generated, and the robots then spend inordinate amounts of time trying to drive through walls. We updated our method to add a check to see if the new map position has jumped a far distance from the prior map position, and if so, we threw that data out.

Image: Michigan Tech Research Institute

In open spaces like the rooms in the Urban circuit, we adjusted our approach to exploration through graph generation to allow the robots to accurately identify viable routes while helping to prevent forays off platform edges.

Our approach to exploration through graph generation based on identification of interior spaces allowed us to thoroughly explore the centers of rooms, although we did have to make some changes from the Tunnel circuit to achieve that. In the Tunnel circuit, we used a simplified graph of the environment based on landmarks like intersections. The advantage of this approach is that it is straightforward for two robots to compare how the graphs of the space they explored individually overlap. In open spaces like the rooms in the Urban circuit, we chose to instead use a more complex, less directly comparable graph structure based on the individual robot’s trajectory. This allowed the robots to accurately identify viable routes between features like subway station platforms and subway tracks, as well as to build up the navigation space for room interiors, while helping to prevent forays off the platform edges. Frontier information is also integrated into the graph, providing a uniform data structure for both goal selection and route planning.

The results are in!
The award ceremony for the Urban Circuit was held concurrently with the Systems Track competition awards this past February in Washington State. We sent a team representative to participate in the Technical Interchange Meeting and present the approach for our team, and the rest of us followed along from our office space on the DARPAtv live stream. While we were confident in our solution, we had also been tracking the online leaderboard and knew our competitors were going to be submitting strong solutions. Since the competition environments are hand-designed, there are always novel challenges that could be presented in these environments as well. We knew we would put up a good fight, but it was very exciting to see BARCS appear in first place!

Any time we implement a new module in our control system, there is a lot of parameter tuning that has to happen to produce reliably good autonomous behavior. In the Urban Circuit, we did not sufficiently test some parameter values in our exploration modules. The effect of this was that the robots only chose to go down small hallways after they explored everything else in their environment, which meant very often they ran out of time and missed a lot of small rooms. This may be the biggest source of lost points for us in the Urban Circuit. One of our major plans going forward from the Urban Circuit is to integrate more sophisticated node selection methods, which can help our robots more intelligently prioritize which frontier nodes to visit. By going through all three Circuit challenges, we will learn how to appropriately add weights to the frontiers based on features of the individual environments. For the Final Challenge, when all three Circuit environments will be combined into large systems, we plan to implement adaptive controllers that will identify their environments and use the appropriate optimized parameters for that environment. In this way, we expect our agents to be able to (for example) prioritize hallways and other small spaces in Urban environments, and perhaps prioritize large openings over small in the Cave environments, if the small openings end up being treacherous overall.

Next for our team: Cave Circuit
Coming up next for Team BARCS is the Virtual Cave Circuit. We are in the middle of testing our hypothesis that our controller will transition from UGVs to UAVs and developing strategies for refining our solution to handle Cave Circuit environmental hazards. The UAVs have a shorter battery life than the UGVs, so executing a joint exploration strategy will also be a high priority for this event, as will completing our work on graph sharing and merging, which will give our robot teams more sophisticated options for navigation and teamwork. We’re reaching a threshold in development where we can start increasing the “smarts” of the robots, which we anticipate will be critical for the final competition, where all of the challenges of SubT will be combined to push the limits of innovation. The Cave Circuit will also have new environmental challenges to tackle: dynamic features such as rock falls have been added, which will block previously accessible passages in the cave environment. We think our controllers are well-poised to handle this new challenge, and we’re eager to find out if that’s the case.

As of now, the biggest worries for us are time and team composition. The Cave Circuit deadline has been postponed to October 15 due to COVID-19 delays, with the award ceremony in mid-November, but there have also been several very compelling additions to the testbed that we would like to experiment with before submission, including droppable networking ‘breadcrumbs’ and new simulated platforms. There are design trade-offs when balancing general versus specialist approaches to the controllers for these robots—since we are adding UAVs to our team for the first time, there are new decisions that will have to be made. For example, the UAVs can ascend into vertical spaces, but only have a battery life of 20 minutes. The UGVs by contrast have 90 minute battery life. One of our strategies is to do an early return to base with one or more agents to buy down risk on making any artifact reports at all for the run, hedging against our other robots not making it back in time, a lesson learned from the Tunnel Circuit. Should a UAV take on this role, or is it better to have them explore deeper into the environment and instead report their artifacts to a UGV or network node, which comes with its own risks? Testing and experimentation to determine the best options takes time, which is always a worry when preparing for a competition! We also anticipate new competitors and stiffer competition all around.

Image: Michigan Tech Research Institute

Team BARCS has now a year to prepare for the final DARPA SubT Challenge event, expected to take place in late 2021.

Going forward from the Cave Circuit, we will have a year to prepare for the final DARPA SubT Challenge event, expected to take place in late 2021. What we are most excited about is increasing the level of intelligence of the agents in their teamwork and joint exploration of the environment. Since we will have (hopefully) built up robust approaches to handling each of the specific types of environments in the Tunnel, Urban, and Cave circuits, we will be aiming to push the limits on collaboration and efficiency among the agents in our team. We view this as a central research contribution of the Virtual Track to the Subterranean Challenge because intelligent, adaptive, multi-robot collaboration is an upcoming stage of development for integration of robots into our lives.

The Subterranean Challenge Virtual Track gives us a bridge for transitioning our more abstract research ideas and algorithms relevant to this degree of autonomy and collaboration onto physical systems, and exploring the tangible outcomes of implementing our work in the real world. And the next time there’s an incident in the basement of our building, the robots (and humans) of Team BARCS will be ready to respond.

Richard Chase, Ph.D., P.E., is a research scientist at Michigan Tech Research Institute (MTRI) and has 20 years of experience developing robotics and cyber physical systems in areas from remote sensing to autonomous vehicles. At MTRI, he works on a variety of topics such as swarm autonomy, human-swarm teaming, and autonomous vehicles. His research interests are the intersection of design, robotics, and embedded systems.

Sarah Kitchen is a Ph.D. mathematician working as a research scientist and an AI/Robotics focus area leader at MTRI. Her research interests include intelligent autonomous agents and multi-agent collaborative teams, as well as applications of autonomous robots to sensing systems.

This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Contract No. HR001118C0124 and is released under Distribution Statement (Approved for Public Release, Distribution Unlimited). Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of DARPA. Continue reading →

Posted in Human Robots

#437471 How Giving Robots a Hybrid, Human-Like ...

Posted on October 26, 2020 by Android

Squeezing a lot of computing power into robots without using up too much space or energy is a constant battle for their designers. But a new approach that mimics the structure of the human brain could provide a workaround.

The capabilities of most of today’s mobile robots are fairly rudimentary, but giving them the smarts to do their jobs is still a serious challenge. Controlling a body in a dynamic environment takes a surprising amount of processing power, which requires both real estate for chips and considerable amounts of energy to power them.

As robots get more complex and capable, those demands are only going to increase. Today’s most powerful AI systems run in massive data centers across far more chips than can realistically fit inside a machine on the move. And the slow death of Moore’s Law suggests we can’t rely on conventional processors getting significantly more efficient or compact anytime soon.

That prompted a team from the University of Southern California to resurrect an idea from more than 40 years ago: mimicking the human brain’s division of labor between two complimentary structures. While the cerebrum is responsible for higher cognitive functions like vision, hearing, and thinking, the cerebellum integrates sensory data and governs movement, balance, and posture.

When the idea was first proposed the technology didn’t exist to make it a reality, but in a paper recently published in Science Robotics, the researchers describe a hybrid system that combines analog circuits that control motion and digital circuits that govern perception and decision-making in an inverted pendulum robot.

“Through this cooperation of the cerebrum and the cerebellum, the robot can conduct multiple tasks simultaneously with a much shorter latency and lower power consumption,” write the researchers.

The type of robot the researchers were experimenting with looks essentially like a pole balancing on a pair of wheels. They have a broad range of applications, from hoverboards to warehouse logistics—Boston Dynamics’ recently-unveiled Handle robot operates on the same principles. Keeping them stable is notoriously tough, but the new approach managed to significantly improve all digital control approaches by radically improving the speed and efficiency of computations.

Key to bringing the idea alive was the recent emergence of memristors—electrical components whose resistance relies on previous input, which allows them to combine computing and memory in one place in a way similar to how biological neurons operate.

The researchers used memristors to build an analog circuit that runs an algorithm responsible for integrating data from the robot’s accelerometer and gyroscope, which is crucial for detecting the angle and velocity of its body, and another that controls its motion. One key advantage of this setup is that the signals from the sensors are analog, so it does away with the need for extra circuitry to convert them into digital signals, saving both space and power.

More importantly, though, the analog system is an order of magnitude faster and more energy-efficient than a standard all-digital system, the authors report. This not only lets them slash the power requirements, but also lets them cut the processing loop from 3,000 microseconds to just 6. That significantly improves the robot’s stability, with it taking just one second to settle into a steady state compared to more than three seconds using the digital-only platform.

At the minute this is just a proof of concept. The robot the researchers have built is small and rudimentary, and the algorithms being run on the analog circuit are fairly basic. But the principle is a promising one, and there is currently a huge amount of R&D going into neuromorphic and memristor-based analog computing hardware.

As often turns out to be the case, it seems like we can’t go too far wrong by mimicking the best model of computation we have found so far: our own brains.

Image Credit: Photos Hobby / Unsplash Continue reading →

Posted in Human Robots

#437395 Microsoft Had a Crazy Idea to Put ...

Posted on September 18, 2020 by Android

A little over two years ago, a shipping container-sized cylinder bearing Microsoft’s name and logo was lowered onto the ocean floor off the northern coast of Scotland. Inside were 864 servers, and their submersion was part of the second phase of the software giant’s Project Natick. Launched in 2015, the project’s purpose is to determine the feasibility of underwater data centers powered by offshore renewable energy.

A couple months ago, the deep-sea servers were brought back up to the surface so engineers could inspect them and evaluate how they’d performed while under water.

But wait—why were they there in the first place?

As bizarre as it seems to sink hundreds of servers into the ocean, there are actually several very good reasons to do so. According to the UN, about 40 percent of the world’s population lives within 60 miles of an ocean. As internet connectivity expands to cover most of the globe in the next few years, millions more people will come online, and a lot more servers will be needed to manage the increased demand and data they’ll generate.

In densely-populated cities real estate is expensive and can be hard to find. But know where there’s lots of cheap, empty space? At the bottom of the ocean. This locale also carries the added benefit of being really cold (depending where we’re talking, that is; if you’re looking off the coast of, say, Mumbai or Abu Dhabi, the waters are warmer).

Servers generate a lot of heat, and datacenters use most of their electricity for cooling. Keeping not just the temperature but also the humidity level constant is important for optimal functioning of the servers; neither of these vary much 100 feet under water.

Finally, installing data centers on the ocean floor is, surprisingly, much faster than building them on land. Microsoft claims its server-holding cylinders will take less than 90 days to go from factory ship to operation, as compared to the average two years it takes to get a terrestrial data center up and running.

Microsoft’s Special Projects team operated the underwater data center for two years, and it took a full day to dredge it up and bring it to the surface. One of the first things researchers did was to insert test tubes into the container to take samples of the air inside; they’ll use it to try to determine how gases released from the equipment may have impacted the servers’ operating environment.

The container was filled with dry nitrogen upon deployment, which seems to have made for a much better environment than the oxygen that land-bound servers are normally surrounded by; the failure rate of the servers in the water was just one-eighth that of Microsoft’s typical rate for its servers on land. The team thinks the nitrogen atmosphere was helpful because it’s less corrosive than oxygen. The fact that no humans entered the container for the entirety of its operations helped, too (no moving around of components or having to turn on lights or adjust the temperature).

Ben Cutler, a project manager in Microsoft’s Special Projects research group who leads Project Natick, believes the results of this phase of the project are sufficient to show that underwater data centers are worth pursuing. “We are now at the point of trying to harness what we have done as opposed to feeling the need to go and prove out some more,” he said.

Cutler envisions putting underwater datacenters near offshore wind farms to power them sustainably. The data centers of the future will require less human involvement, instead being managed and run primarily by technologies like robotics and AI. In this kind of “lights-out” datacenter, the servers would be swapped out about once every five years, with any that fail before then being taken offline.

The final step in this phase of Project Natick is to recycle all the components used for the underwater data center, including the steel pressure vessel, heat exchangers, and the servers themselves—and restoring the sea bed where the cylinder rested back to its original condition.

If Cutler’s optimism is a portent of things to come, it may not be long before the ocean floor is dotted with sustainable datacenters to feed our ever-increasing reliance on our phones and the internet.

Image Credit: Microsoft Continue reading →

Posted in Human Robots

#437303 The Deck Is Not Rigged: Poker and the ...

Posted on August 8, 2020 by Android

Tuomas Sandholm, a computer scientist at Carnegie Mellon University, is not a poker player—or much of a poker fan, in fact—but he is fascinated by the game for much the same reason as the great game theorist John von Neumann before him. Von Neumann, who died in 1957, viewed poker as the perfect model for human decision making, for finding the balance between skill and chance that accompanies our every choice. He saw poker as the ultimate strategic challenge, combining as it does not just the mathematical elements of a game like chess but the uniquely human, psychological angles that are more difficult to model precisely—a view shared years later by Sandholm in his research with artificial intelligence.

“Poker is the main benchmark and challenge program for games of imperfect information,” Sandholm told me on a warm spring afternoon in 2018, when we met in his offices in Pittsburgh. The game, it turns out, has become the gold standard for developing artificial intelligence.

Tall and thin, with wire-frame glasses and neat brow hair framing a friendly face, Sandholm is behind the creation of three computer programs designed to test their mettle against human poker players: Claudico, Libratus, and most recently, Pluribus. (When we met, Libratus was still a toddler and Pluribus didn’t yet exist.) The goal isn’t to solve poker, as such, but to create algorithms whose decision making prowess in poker’s world of imperfect information and stochastic situations—situations that are randomly determined and unable to be predicted—can then be applied to other stochastic realms, like the military, business, government, cybersecurity, even health care.

While the first program, Claudico, was summarily beaten by human poker players—“one broke-ass robot,” an observer called it—Libratus has triumphed in a series of one-on-one, or heads-up, matches against some of the best online players in the United States.

Libratus relies on three main modules. The first involves a basic blueprint strategy for the whole game, allowing it to reach a much faster equilibrium than its predecessor. It includes an algorithm called the Monte Carlo Counterfactual Regret Minimization, which evaluates all future actions to figure out which one would cause the least amount of regret. Regret, of course, is a human emotion. Regret for a computer simply means realizing that an action that wasn’t chosen would have yielded a better outcome than one that was. “Intuitively, regret represents how much the AI regrets having not chosen that action in the past,” says Sandholm. The higher the regret, the higher the chance of choosing that action next time.

It’s a useful way of thinking—but one that is incredibly difficult for the human mind to implement. We are notoriously bad at anticipating our future emotions. How much will we regret doing something? How much will we regret not doing something else? For us, it’s an emotionally laden calculus, and we typically fail to apply it in quite the right way. For a computer, it’s all about the computation of values. What does it regret not doing the most, the thing that would have yielded the highest possible expected value?

The second module is a sub-game solver that takes into account the mistakes the opponent has made so far and accounts for every hand she could possibly have. And finally, there is a self-improver. This is the area where data and machine learning come into play. It’s dangerous to try to exploit your opponent—it opens you up to the risk that you’ll get exploited right back, especially if you’re a computer program and your opponent is human. So instead of attempting to do that, the self-improver lets the opponent’s actions inform the areas where the program should focus. “That lets the opponent’s actions tell us where [they] think they’ve found holes in our strategy,” Sandholm explained. This allows the algorithm to develop a blueprint strategy to patch those holes.

It’s a very human-like adaptation, if you think about it. I’m not going to try to outmaneuver you head on. Instead, I’m going to see how you’re trying to outmaneuver me and respond accordingly. Sun-Tzu would surely approve. Watch how you’re perceived, not how you perceive yourself—because in the end, you’re playing against those who are doing the perceiving, and their opinion, right or not, is the only one that matters when you craft your strategy. Overnight, the algorithm patches up its overall approach according to the resulting analysis.

There’s one final thing Libratus is able to do: play in situations with unknown probabilities. There’s a concept in game theory known as the trembling hand: There are branches of the game tree that, under an optimal strategy, one should theoretically never get to; but with some probability, your all-too-human opponent’s hand trembles, they take a wrong action, and you’re suddenly in a totally unmapped part of the game. Before, that would spell disaster for the computer: An unmapped part of the tree means the program no longer knows how to respond. Now, there’s a contingency plan.

Of course, no algorithm is perfect. When Libratus is playing poker, it’s essentially working in a zero-sum environment. It wins, the opponent loses. The opponent wins, it loses. But while some real-life interactions really are zero-sum—cyber warfare comes to mind—many others are not nearly as straightforward: My win does not necessarily mean your loss. The pie is not fixed, and our interactions may be more positive-sum than not.

What’s more, real-life applications have to contend with something that a poker algorithm does not: the weights that are assigned to different elements of a decision. In poker, this is a simple value-maximizing process. But what is value in the human realm? Sandholm had to contend with this before, when he helped craft the world’s first kidney exchange. Do you want to be more efficient, giving the maximum number of kidneys as quickly as possible—or more fair, which may come at a cost to efficiency? Do you want as many lives as possible saved—or do some take priority at the cost of reaching more? Is there a preference for the length of the wait until a transplant? Do kids get preference? And on and on. It’s essential, Sandholm says, to separate means and the ends. To figure out the ends, a human has to decide what the goal is.

“The world will ultimately become a lot safer with the help of algorithms like Libratus,” Sandholm told me. I wasn’t sure what he meant. The last thing that most people would do is call poker, with its competition, its winners and losers, its quest to gain the maximum edge over your opponent, a haven of safety.

“Logic is good, and the AI is much better at strategic reasoning than humans can ever be,” he explained. “It’s taking out irrationality, emotionality. And it’s fairer. If you have an AI on your side, it can lift non-experts to the level of experts. Naïve negotiators will suddenly have a better weapon. We can start to close off the digital divide.”

It was an optimistic note to end on—a zero-sum, competitive game yielding a more ultimately fair and rational world.

I wanted to learn more, to see if it was really possible that mathematics and algorithms could ultimately be the future of more human, more psychological interactions. And so, later that day, I accompanied Nick Nystrom, the chief scientist of the Pittsburgh Supercomputing Center—the place that runs all of Sandholm’s poker-AI programs—to the actual processing center that make undertakings like Libratus possible.

A half-hour drive found us in a parking lot by a large glass building. I’d expected something more futuristic, not the same square, corporate glass squares I’ve seen countless times before. The inside, however, was more promising. First the security checkpoint. Then the ride in the elevator — down, not up, to roughly three stories below ground, where we found ourselves in a maze of corridors with card readers at every juncture to make sure you don’t slip through undetected. A red-lit panel formed the final barrier, leading to a small sliver of space between two sets of doors. I could hear a loud hum coming from the far side.

“Let me tell you what you’re going to see before we walk in,” Nystrom told me. “Once we get inside, it will be too loud to hear.”

I was about to witness the heart of the supercomputing center: 27 large containers, in neat rows, each housing multiple processors with speeds and abilities too great for my mind to wrap around. Inside, the temperature is by turns arctic and tropic, so-called “cold” rows alternating with “hot”—fans operate around the clock to cool the processors as they churn through millions of giga, mega, tera, peta and other ever-increasing scales of data bytes. In the cool rows, robotic-looking lights blink green and blue in orderly progression. In the hot rows, a jumble of multicolored wires crisscrosses in tangled skeins.

In the corners stood machines that had outlived their heyday. There was Sherlock, an old Cray model, that warmed my heart. There was a sad nameless computer, whose anonymity was partially compensated for by the Warhol soup cans adorning its cage (an homage to Warhol’s Pittsburghian origins).

And where does Libratus live, I asked? Which of these computers is Bridges, the computer that runs the AI Sandholm and I had been discussing?

Bridges, it turned out, isn’t a single computer. It’s a system with processing power beyond comprehension. It takes over two and a half petabytes to run Libratus. A single petabyte is a million gigabytes: You could watch over 13 years of HD video, store 10 billion photos, catalog the contents of the entire Library of Congress word for word. That’s a whole lot of computing power. And that’s only to succeed at heads-up poker, in limited circumstances.

Yet despite the breathtaking computing power at its disposal, Libratus is still severely limited. Yes, it beat its opponents where Claudico failed. But the poker professionals weren’t allowed to use many of the tools of their trade, including the opponent analysis software that they depend on in actual online games. And humans tire. Libratus can churn for a two-week marathon, where the human mind falters.

But there’s still much it can’t do: play more opponents, play live, or win every time. There’s more humanity in poker than Libratus has yet conquered. “There’s this belief that it’s all about statistics and correlations. And we actually don’t believe that,” Nystrom explained as we left Bridges behind. “Once in a while correlations are good, but in general, they can also be really misleading.”

Two years later, the Sandholm lab will produce Pluribus. Pluribus will be able to play against five players—and will run on a single computer. Much of the human edge will have evaporated in a short, very short time. The algorithms have improved, as have the computers. AI, it seems, has gained by leaps and bounds.

So does that mean that, ultimately, the algorithmic can indeed beat out the human, that computation can untangle the web of human interaction by discerning “the little tactics of deception, of asking yourself what is the other man going to think I mean to do,” as von Neumann put it?

Long before I’d spoken to Sandholm, I’d met Kevin Slavin, a polymath of sorts whose past careers have including founding a game design company and an interactive art space and launching the Playful Systems group at MIT’s Media Lab. Slavin has a decidedly different view from the creators of Pluribus. “On the one hand, [von Neumann] was a genius,” Kevin Slavin reflects. “But the presumptuousness of it.”

Slavin is firmly on the side of the gambler, who recognizes uncertainty for what it is and thus is able to take calculated risks when necessary, all the while tampering confidence at the outcome. The most you can do is put yourself in the path of luck—but to think you can guess with certainty the actual outcome is a presumptuousness the true poker player foregoes. For Slavin, the wonder of computers is “That they can generate this fabulous, complex randomness.” His opinion of the algorithmic assaults on chance? “This is their moment,” he said. “But it’s the exact opposite of what’s really beautiful about a computer, which is that it can do something that’s actually unpredictable. That, to me, is the magic.”

Will they actually succeed in making the unpredictable predictable, though? That’s what I want to know. Because everything I’ve seen tells me that absolute success is impossible. The deck is not rigged.

“It’s an unbelievable amount of work to get there. What do you get at the end? Let’s say they’re successful. Then we live in a world where there’s no God, agency, or luck,” Slavin responded.

“I don’t want to live there,’’ he added “I just don’t want to live there.”

Luckily, it seems that for now, he won’t have to. There are more things in life than are yet written in the algorithms. We have no reliable lie detection software—whether in the face, the skin, or the brain. In a recent test of bluffing in poker, computer face recognition failed miserably. We can get at discomfort, but we can’t get at the reasons for that discomfort: lying, fatigue, stress—they all look much the same. And humans, of course, can also mimic stress where none exists, complicating the picture even further.

Pluribus may turn out to be powerful, but von Neumann’s challenge still stands: The true nature of games, the most human of the human, remains to be conquered.

This article was originally published on Undark. Read the original article.

Image Credit: José Pablo Iglesias / Unsplash Continue reading →

Posted in Human Robots

#437269 DeepMind’s Newest AI Programs Itself ...

Posted on July 27, 2020 by Android

When Deep Blue defeated world chess champion Garry Kasparov in 1997, it may have seemed artificial intelligence had finally arrived. A computer had just taken down one of the top chess players of all time. But it wasn’t to be.

Though Deep Blue was meticulously programmed top-to-bottom to play chess, the approach was too labor-intensive, too dependent on clear rules and bounded possibilities to succeed at more complex games, let alone in the real world. The next revolution would take a decade and a half, when vastly more computing power and data revived machine learning, an old idea in artificial intelligence just waiting for the world to catch up.

Today, machine learning dominates, mostly by way of a family of algorithms called deep learning, while symbolic AI, the dominant approach in Deep Blue’s day, has faded into the background.

Key to deep learning’s success is the fact the algorithms basically write themselves. Given some high-level programming and a dataset, they learn from experience. No engineer anticipates every possibility in code. The algorithms just figure it.

Now, Alphabet’s DeepMind is taking this automation further by developing deep learning algorithms that can handle programming tasks which have been, to date, the sole domain of the world’s top computer scientists (and take them years to write).

In a paper recently published on the pre-print server arXiv, a database for research papers that haven’t been peer reviewed yet, the DeepMind team described a new deep reinforcement learning algorithm that was able to discover its own value function—a critical programming rule in deep reinforcement learning—from scratch.

Surprisingly, the algorithm was also effective beyond the simple environments it trained in, going on to play Atari games—a different, more complicated task—at a level that was, at times, competitive with human-designed algorithms and achieving superhuman levels of play in 14 games.

DeepMind says the approach could accelerate the development of reinforcement learning algorithms and even lead to a shift in focus, where instead of spending years writing the algorithms themselves, researchers work to perfect the environments in which they train.

Pavlov’s Digital Dog
First, a little background.

Three main deep learning approaches are supervised, unsupervised, and reinforcement learning.

The first two consume huge amounts of data (like images or articles), look for patterns in the data, and use those patterns to inform actions (like identifying an image of a cat). To us, this is a pretty alien way to learn about the world. Not only would it be mind-numbingly dull to review millions of cat images, it’d take us years or more to do what these programs do in hours or days. And of course, we can learn what a cat looks like from just a few examples. So why bother?

While supervised and unsupervised deep learning emphasize the machine in machine learning, reinforcement learning is a bit more biological. It actually is the way we learn. Confronted with several possible actions, we predict which will be most rewarding based on experience—weighing the pleasure of eating a chocolate chip cookie against avoiding a cavity and trip to the dentist.

In deep reinforcement learning, algorithms go through a similar process as they take action. In the Atari game Breakout, for instance, a player guides a paddle to bounce a ball at a ceiling of bricks, trying to break as many as possible. When playing Breakout, should an algorithm move the paddle left or right? To decide, it runs a projection—this is the value function—of which direction will maximize the total points, or rewards, it can earn.

Move by move, game by game, an algorithm combines experience and value function to learn which actions bring greater rewards and improves its play, until eventually, it becomes an uncanny Breakout player.

Learning to Learn (Very Meta)
So, a key to deep reinforcement learning is developing a good value function. And that’s difficult. According to the DeepMind team, it takes years of manual research to write the rules guiding algorithmic actions—which is why automating the process is so alluring. Their new Learned Policy Gradient (LPG) algorithm makes solid progress in that direction.

LPG trained in a number of toy environments. Most of these were “gridworlds”—literally two-dimensional grids with objects in some squares. The AI moves square to square and earns points or punishments as it encounters objects. The grids vary in size, and the distribution of objects is either set or random. The training environments offer opportunities to learn fundamental lessons for reinforcement learning algorithms.

Only in LPG’s case, it had no value function to guide that learning.

Instead, LPG has what DeepMind calls a “meta-learner.” You might think of this as an algorithm within an algorithm that, by interacting with its environment, discovers both “what to predict,” thereby forming its version of a value function, and “how to learn from it,” applying its newly discovered value function to each decision it makes in the future.

Prior work in the area has had some success, but according to DeepMind, LPG is the first algorithm to discover reinforcement learning rules from scratch and to generalize beyond training. The latter was particularly surprising because Atari games are so different from the simple worlds LPG trained in—that is, it had never seen anything like an Atari game.

Time to Hand Over the Reins? Not Just Yet
LPG is still behind advanced human-designed algorithms, the researchers said. But it outperformed a human-designed benchmark in training and even some Atari games, which suggests it isn’t strictly worse, just that it specializes in some environments.

This is where there’s room for improvement and more research.

The more environments LPG saw, the more it could successfully generalize. Intriguingly, the researchers speculate that with enough well-designed training environments, the approach might yield a general-purpose reinforcement learning algorithm.

At the least, though, they say further automation of algorithm discovery—that is, algorithms learning to learn—will accelerate the field. In the near term, it can help researchers more quickly develop hand-designed algorithms. Further out, as self-discovered algorithms like LPG improve, engineers may shift from manually developing the algorithms themselves to building the environments where they learn.

Deep learning long ago left Deep Blue in the dust at games. Perhaps algorithms learning to learn will be a winning strategy in the real world too.

Image credit: Mike Szczepanski / Unsplash Continue reading →

Posted in Human Robots