Tag Archives: player

#438695 Technology based on rodent neurons may ...

Imagine you are playing an immersive game in which you are dropped into an unknown landscape with a directive to find a certain location. To advance forward in the game, you must also map the terrain so that you can then share your initial location and your map with another remote player. You have now been given a problem that, within the world of robotics is called SLAM. You have been asked to simultaneously localize and map an unknown environment. Continue reading

Posted in Human Robots

#437721 Video Friday: Child Robot Learning to ...

Video Friday is your weekly selection of awesome robotics videos, collected by your Automaton bloggers. We’ll also be posting a weekly calendar of upcoming robotics events for the next few months; here’s what we have so far (send us your events!):

CLAWAR 2020 – August 24-26, 2020 – [Online Conference]
ICUAS 2020 – September 1-4, 2020 – Athens, Greece
ICRES 2020 – September 28-29, 2020 – Taipei, Taiwan
AUVSI EXPONENTIAL 2020 – October 5-8, 2020 – [Online Conference]
IROS 2020 – October 25-29, 2020 – Las Vegas, Nev., USA
CYBATHLON 2020 – November 13-14, 2020 – [Online Event]
ICSR 2020 – November 14-16, 2020 – Golden, Colo., USA
Let us know if you have suggestions for next week, and enjoy today’s videos.

We first met Ibuki, Hiroshi Ishiguro’s latest humanoid robot, a couple of years ago. A recent video shows how Ishiguro and his team are teaching the robot to express its emotional state through gait and body posture while moving.

This paper presents a subjective evaluation of the emotions of a wheeled mobile humanoid robot expressing emotions during movement by replicating human gait-induced upper body motion. For this purpose, we proposed the robot equipped with a vertical oscillation mechanism that generates such motion by focusing on human center-of-mass trajectory. In the experiment, participants watched videos of the robot’s different emotional gait-induced upper body motions, and assess the type of emotion shown, and their confidence level in their answer.

[ Hiroshi Ishiguro Lab ] via [ RobotStart ]

ICYMI: This is a zinc-air battery made partly of Kevlar that can be used to support weight, not just add to it.

Like biological fat reserves store energy in animals, a new rechargeable zinc battery integrates into the structure of a robot to provide much more energy, a team led by the University of Michigan has shown.

The new battery works by passing hydroxide ions between a zinc electrode and the air side through an electrolyte membrane. That membrane is partly a network of aramid nanofibers—the carbon-based fibers found in Kevlar vests—and a new water-based polymer gel. The gel helps shuttle the hydroxide ions between the electrodes. Made with cheap, abundant and largely nontoxic materials, the battery is more environmentally friendly than those currently in use. The gel and aramid nanofibers will not catch fire if the battery is damaged, unlike the flammable electrolyte in lithium ion batteries. The aramid nanofibers could be upcycled from retired body armor.

[ University of Michigan ]

In what they say is the first large-scale study of the interactions between sound and robotic action, researchers at CMU’s Robotics Institute found that sounds could help a robot differentiate between objects, such as a metal screwdriver and a metal wrench. Hearing also could help robots determine what type of action caused a sound and help them use sounds to predict the physical properties of new objects.

[ CMU ]

Captured on Aug. 11 during the second rehearsal of the OSIRIS-REx mission’s sample collection event, this series of images shows the SamCam imager’s field of view as the NASA spacecraft approaches asteroid Bennu’s surface. The rehearsal brought the spacecraft through the first three maneuvers of the sampling sequence to a point approximately 131 feet (40 meters) above the surface, after which the spacecraft performed a back-away burn.

These images were captured over a 13.5-minute period. The imaging sequence begins at approximately 420 feet (128 meters) above the surface – before the spacecraft executes the “Checkpoint” maneuver – and runs through to the “Matchpoint” maneuver, with the last image taken approximately 144 feet (44 meters) above the surface of Bennu.

[ NASA ]

The DARPA AlphaDogfight Trials Final Event took place yesterday; the livestream is like 5 hours long, but you can skip ahead to 4:39 ish to see the AI winner take on a human F-16 pilot in simulation.

Some things to keep in mind about the result: The AI had perfect situational knowledge while the human pilot had to use eyeballs, and in particular, the AI did very well at lining up its (virtual) gun with the human during fast passing maneuvers, which is the sort of thing that autonomous systems excel at but is not necessarily reflective of better strategy.

[ DARPA ]

Coming soon from Clearpath Robotics!

[ Clearpath ]

This video introduces Preferred Networks’ Hand type A, a tendon-driven robot gripper with passively switchable underactuated surface.

[ Preferred Networks ]

CYBATHLON 2020 will take place on 13 – 14 November 2020 – at the teams’ home bases. They will set up their infrastructure for the competition and film their races. Instead of starting directly next to each other, the pilots will start individually and under the supervision of CYBATHLON officials. From Zurich, the competitions will be broadcast through a new platform in a unique live programme.

[ Cybathlon ]

In this project, we consider the task of autonomous car racing in the top-selling car racing game Gran Turismo Sport. Gran Turismo Sport is known for its detailed physics simulation of various cars and tracks. Our approach makes use of maximum-entropy deep reinforcement learning and a new reward design to train a sensorimotor policy to complete a given race track as fast as possible. We evaluate our approach in three different time trial settings with different cars and tracks. Our results show that the obtained controllers not only beat the built-in non-player character of Gran Turismo Sport, but also outperform the fastest known times in a dataset of personal best lap times of over 50,000 human drivers.

[ UZH ]

With the help of the software pitasc from Fraunhofer IPA, an assembly task is no longer programmed point by point, but workpiece-related. Thus, pitasc adapts the assembly process itself for new product variants with the help of updated parameters.

[ Fraunhofer ]

In this video, a multi-material robot simulator is used to design a shape-changing robot, which is then transferred to physical hardware. The simulated and real robots can use shape change to switch between rolling gaits and inchworm gaits, to locomote in multiple environments.

[ Yale ]

This work presents a novel loco-manipulation control framework for the execution of complex tasks with kinodynamic constraints using mobile manipulators. As a representative example, we consider the handling and re-positioning of pallet jacks in unstructured environments. While these results reveal with a proof-of- concept the effectiveness of the proposed framework, they also demonstrate the high potential of mobile manipulators for relieving human workers from such repetitive and labor intensive tasks. We believe that this extended functionality can contribute to increasing the usability of mobile manipulators in different application scenarios.

[ Paper ] via [ IIT ]

I don’t know why this dinosaur ice cream serving robot needs to blow smoke out of its nose, but I like it.

[ Connected Robotics ] via [ RobotStart ]

Guardian S remote visual inspection and surveillance robots make laying cable runs in confined or hard to reach spaces easy. With advanced maneuverability and the ability to climb vertical, ferrous surfaces, the robot reaches areas that are not always easily accessible.

[ Sarcos ]

Looks like the company that bought Anki is working on an add-on to let cars charge while they drive.

[ Digital Dream Labs ]

Chris Atkeson gives a brief talk for the CMU Robotics Institute orientation.

[ CMU RI ]

A UofT Robotics Seminar, featuring Russ Tedrake from MIT and TRI on “Feedback Control for Manipulation.”

Control theory has an answer for just about everything, but seems to fall short when it comes to closing a feedback loop using a camera, dealing with the dynamics of contact, and reasoning about robustness over the distribution of tasks one might find in the kitchen. Recent examples from RL and imitation learning demonstrate great promise, but don’t leverage the rigorous tools from systems theory. I’d like to discuss why, and describe some recent results of closing feedback loops from pixels for “category-level” robot manipulation.

[ UofT ] Continue reading

Posted in Human Robots

#437550 McDonald’s Is Making a Plant-Based ...

Fast-food chains have been doing what they can in recent years to health-ify their menus. For better or worse, burgers, fries, fried chicken, roast beef sandwiches, and the like will never go out of style—this is America, after all—but consumers are increasingly gravitating towards healthier options.

One of those options is plant-based foods, and not just salads and veggie burgers, but “meat” made from plants. Burger King was one of the first big fast-food chains to jump on the plant-based meat bandwagon, introducing its Impossible Whopper in restaurants across the country last year after a successful pilot program. Dunkin’ (formerly Dunkin’ Donuts) uses plant-based patties in its Beyond Sausage breakfast sandwiches.

But there’s one big player in the fast food market that’s been oddly missing from the plant-based trend—until now. McDonald’s announced last week that it will debut a sandwich called the McPlant in key US markets next year. Unlike Dunkin’ and Burger King, who both worked with Impossible Foods to make their plant-based products, McDonald’s worked with Los Angeles-based Beyond Meat, which makes chicken, beef, and pork-like products from plants.

According to Bloomberg, though, McDonald’s decided to forego a partnership with Beyond Meat in favor of creating its own plant-based products. Imitation chicken nuggets and plant-based breakfast sandwiches are in its plans as well.

McDonald’s has bounced back impressively from its March low (when the coronavirus lockdowns first happened in the US). Last month the company’s stock reached a 52-week high of $231 per share (as compared to its low in March of $124 per share).

To keep those numbers high and make it as easy as possible for customers to get their hands on plant-based burgers and all the traditional menu items too, the fast food chain is investing in tech and integrating more digital offerings into its restaurants.

McDonald’s has acquired a couple artificial intelligence companies in the last year and a half; Dynamic Yield is an Israeli company that uses AI to personalize customers’ experiences, and McDonald’s is using Dynamic Yield’s tech on its smart menu boards, for example by customizing the items displayed on the drive-thru menu based on the weather and the time of day, and recommending additional items based on what a customer asks for first (i.e. “You know what would go great with that coffee? Some pancakes!”).

The fast food giant also bought Apprente, a startup that uses AI in voice-based ordering platforms. McDonald’s is using the tech to help automate its drive-throughs.

In addition to these investments, the company plans to launch a digital hub called MyMcDonald’s that will include a loyalty program, start doing deliveries of its food through its mobile app, and test different ways of streamlining the food order and pickup process—with many of the new ideas geared towards pandemic times, like express pickup lanes for people who placed digital orders and restaurants with drive-throughs for delivery and pickup orders only.

Plant-based meat patties appear to be just one small piece of McDonald’s modernization plans. Those of us who were wondering what they were waiting for should have known—one of the most-recognized fast food chains in the world wasn’t about to let itself get phased out. It seems it will only be a matter of time until you can pull out your phone, make a few selections, and have a burger made from plants—with a side of fries made from more plants—show up at your door a little while later. Drive-throughs, shouting your order into a fuzzy speaker with a confused teen on the other end, and burgers made from beef? So 2019.

Image Credit: McDonald’s Continue reading

Posted in Human Robots

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

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 ...

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