Full width home advertisement

Welcome Home


Post Page Advertisement [Top]

Even though artificial intelligence is intelligent, it does not interact well with others



According to a study, when humans play cooperative games together, they find artificial intelligence to be a frustrating teammate, posing challenges for "teaming intelligence."


Using artificial intelligence (AI) programs, computer programs have beaten the world's best players in games such as chess and Go by a wide margin. However, while these "superhuman" artificial intelligences are unbeatable competitors, collaborating with them may be more difficult than competing against people. How long do you think it will take for one technology to coexist with another?


Scientists at the Massachusetts Institute of Technology Lincoln Laboratory conducted a new study to determine how well humans could play the cooperative card game Hanabi when paired with an advanced artificial intelligence model trained to excel at playing with teammates it had never met before. A series of games was played by participants in single-blind experiments: one with an AI agent as their teammate, and another with a rule-based agent, which was programmed to behave in a predefined way.


The findings took the researchers completely by surprise. Apart from the fact that humans consistently expressed displeasure with their AI teammate, the scores with the AI teammate were the same as with the rule-based agent. Their reactions were characterized as unpredictable, unreliable, and untrustworthy, and they felt negative emotions even when the team performed admirably. A paper describing this research has been accepted for presentation at the 2021 Conference on Neural Information Processing Systems (NeurIPS).


In the words of Ross Allen, co-author of the paper and a researcher in the Artificial Intelligence Technology Group, "It exemplifies the subtle distinction between developing AI that performs well objectively and developing AI that is subjectively trusted or preferred." "These two issues may appear to be inextricably linked, but according to this study they are in fact two separate issues with no connection. We need to figure out how to untangle them."


It is possible that humans will despise their artificial intelligence teammates in the future, which could pose a problem for researchers who are developing AI technology to work alongside humans on real-world challenges, such as defending against missiles or performing complex surgery. It is this dynamic, which has been dubbed teaming intelligence, that represents a new frontier in artificial intelligence research, as it makes use of a type of AI known as reinforcement learning.


It is not necessary to tell a reinforcement learning AI which actions to perform; rather, it learns which actions result in the greatest numerical "reward" by repeatedly testing different scenarios. In the course of research, superhuman chess and Go players have been created as a result of this technology. These artificial intelligence systems, in contrast to rule-based algorithms, are not programmed with "if/then" statements because the possible outcomes of the human tasks they are expected to perform, such as driving a car, are far too numerous to code.


"Reinforcement learning is a method of developing artificial intelligence that is significantly more general-purpose in nature. Even if you can teach an agent to play chess, it is unlikely that he or she will go out and drive a car. But with the right data, you can use the same algorithms to train a different agent to drive a car, for example "Allen explains further. As far as what it could theoretically accomplish is concerned, "the sky is the limit."


Researchers are now using Hanabi to evaluate the performance of collaborative reinforcement learning models, in a manner similar to how chess has been used for decades to do the same.


Hanabi is a card game that plays similarly to a multiplayer version of Solitaire. Using their collective efforts, the players stack cards of the same suit in the correct order. Players, on the other hand, are not permitted to see their own cards; instead, they are only permitted to see those of their teammates. Each player has a strict limit on their ability to communicate with their teammates in order to persuade them to stack the best card from their own hand as the next card in the stack.


There is no evidence that the artificial intelligence or rule-based agents used in this experiment were developed by the researchers at Lincoln Laboratory. Regarding Hanabi performance, both agents are among the best in their respective fields, according to Hanabi. According to these findings, when the AI model had previously been paired with an unidentified AI teammate, the team received the highest possible score for Hanabi play between two unknown AI agents.


"It was a significant outcome," Allen says of the outcome. "After all, if these AIs have never met before and can play well together, we should be able to bring humans who know how to play well together with the AI and expect them to perform well as well as they do themselves. The reason we believed the AI team would perform objectively better was that we believed humans would prefer it, since humans are generally drawn to well-performing objects."


It turned out that neither of those predictions came true. Neither the AI nor the rule-based agent achieved statistically significant differences in performance. In surveys, all 29 participants expressed a strong preference for the rule-based teammate, indicating that they had a strong preference for them. The participants were not aware of which agent had been assigned to them for which games during the competition.


In the paper, Jaime Pena, a researcher in the AI Technology and Systems Group and co-author of the paper, explains that one participant reported that they were so stressed out by the AI agent's poor play that they developed a headache as a result of the agent's poor performance. "According to another, the rule-based agent was stupid but able to complete the task, whereas the AI agent demonstrated that it understood the rules but made decisions that were inconsistent with how a team functions. Poor hints and poor plays, according to them, were the cause of the situation."


As previously observed in reinforcement learning research, this perception of AI making "bad plays" corresponds to unexpected behavior observed in the field. If we look at the case of DeepMind's AlphaGo, which defeated one of the world's best Go players for the first time in 2016, one of the most widely praised moves made by AlphaGo was move 37 in game 2, which was so unusual that human commentators mistook it for a typo. Later analysis revealed that the maneuver had been extremely well-calculated, earning it the title of "genius."


While such moves may be applauded when performed by an artificial intelligence opponent, they are less likely to be applauded when performed by a human team. In these closely coupled teams, the Lincoln Laboratory researchers discovered that unusual or seemingly illogical actions were the primary culprits in eroding humans' trust in their AI teammate, according to their findings. When players perceive that they and their AI teammate are collaborating effectively, they are less likely to want to work with the AI in the future. This is especially true when any potential payoff is not immediately apparent.


A lot of people expressed frustration with the system, saying things like 'I despise working with this thing,'" says Hosea Siu, another co-author of the paper and a researcher in the Control and Autonomous Systems Engineering Group.


Participants who self-identified as Hanabi experts, as the vast majority of players in this study did, were more likely to abandon the AI player than other participants. Siu believes that this is a source of concern for AI developers because the technology's primary users will almost certainly be subject matter experts in their respective fields.


"Consider the following scenario: you are training a super-smart artificial intelligence guidance assistant for use in a missile defense scenario. The experts on board your ships, many of whom have been in the industry for more than 25 years, are not trainees; they are experts in their field. If there is an overwhelming expert bias against it in gaming scenarios, then it is almost certain to manifest itself in real-world operations as a result of this bias "he continues.


Researchers stress that the artificial intelligence used in this study was not created with human preferences in mind, as was previously stated. However, this is a contributing factor to the problem — not many are. According to tradition, this collaborative AI model was created to achieve the highest possible score, and its success has been determined by the objective performance of the model as a whole.


Allen asserts that if researchers do not concentrate on subjective human preference, "we will fail to create artificial intelligence that humans actually want to use." "The work on artificial intelligence that improves a very clean number is less complicated to do. It is significantly more difficult to develop artificial intelligence that is effective in this more ambiguous world of human preferences."


The MeRLin (Mission-Ready Reinforcement Learning) project, funded by Lincoln Laboratory's Technology Office and carried out in collaboration with the United States Air Force Artificial Intelligence Accelerator and the Massachusetts Institute of Technology's Department of Electrical Engineering and Computer Science, has as its goal the solution of this more difficult problem. The project is looking into what has prevented collaborative artificial intelligence technology from progressing beyond the realm of games and into more complex reality scenarios.


The ability of the AI to justify its actions, according to the researchers, will help to build trust. This will be the primary focus of their efforts for the upcoming year.


"Consider the following scenario: we reran the experiment, but afterward — and this is much easier said than done — the human could inquire, 'Why did you do that move, I didn't understand what you were trying to say? Our hypothesis is that if the AI could provide insight into what they believed would happen as a result of their actions, humans would say, "Oh, that's a strange way of thinking about it, but I get it now," and they would trust the AI. Even if we did not make any changes to the AI's underlying decision-making, the outcome would be completely different "Allen explains further.


This type of exchange, similar to a post-game huddle, is frequently what allows humans to develop camaraderie and cooperation as a group of individuals.


"Perhaps there is also a bias in the hiring process. The majority of artificial intelligence teams are lacking in individuals who are willing to work on these squishy humans and their soft problems, according to the report "With a chuckle, Siu continues. "People who are interested in mathematics and optimization are the target audience. And that is the starting point, but it is insufficient on its own."


Mastery of a game such as Hanabi between artificial intelligence and humans could open the door to a whole new world of possibilities for collaborating intelligence in the future. In the meantime, until researchers can narrow the gap between how well an AI performs and how well it is liked by humans, technology will most likely remain at the machine versus human level of performance.

1 comment:

Bottom Ad [Post Page]