DeepMind, is a Google-owned division that focuses on developing artificial intelligence-based technologies.
Its London-based AI outfit has created two different types of AI that can use their ‘imagination’ to plan ahead and perform tasks with a higher success rate than AIs without imagination. The researchers argue that giving AI imagination is crucial for them to deal with real-world environments.
The development of this general-purpose AI is dependent on understanding and encoding human abilities that include imagination, curiosity and memory into AI.
These AIs are called Imagination-Augmented Agents, or I2As. These AIs use what DeepMind calls an “imagination encoder” which is essentially a neural network that helps inform the agents’ decision-making process by learning to gather useful relevant data while rejecting irrelevant information.
This helps the AI in deciding what are and what aren't useful predictions about its environment.
In the researchers' papers, the I2A agents were tasked with different situations to test their predictive abilities, “including the puzzle game Sokoban and a spaceship navigation game.”
Putting them to the test, DeepMind used games like Sokoban is a puzzle game in which an alien has to push, but cannot pull, boxes into the right place. To challenge the agent, the researchers created every level algorithmically, and only gave the agent one try to solve it. This way, the agents are encouraged to try different strategies before testing them in real environment, the researchers said.
The agent's demand forward planning and reasoning are evaluated, as well as the the ideal ecosystem in which their capacity for imagination-based planning. DeepMind pitted the I2A agents against imagination-less baselines.
The result shows that the agents with imaginations, performed better that its imagination-less counterparts. With 'imaginations' to predict and think, they can learn to navigate the puzzles with less experience by just extracting information from internal simulations.
The agents worked to try the different strategies internally before carrying them out in real-world scenarios, as DeepMind generated each task level in such a way that the agents may only perform a strategy once.
And when the researchers added a ‘manager’ component that helped them create a plan, the agents can "solve tasks even more efficiently with fewer steps." The planning strategies can even be further enhanced with the addition of various imagination models that have different levels of accuracy.
Another feature is the ability to create plans using a variety of strategies that include opting to carry on with the imagined course of action or choosing to start all over again.
But what described in the researches and papers, these AIs are nowhere near what humans are capable of. They are still imperfect, but they can understand the dynamics of the environment surrounding them. The new types of AI can also adjust the imagined courses of action in order to address problems that are given to them.
Additionally, the imagination encoder improves the AI’s efficiency as it extracts more data from imagination that does not necessarily lead to a high reward.
This certainly show that AI can benefit by having imagination to efficiently plan different scenarios before acting.
Creating agents with an imagination that can rival humans is perhaps the hardest challenge for AI research. Building an agent that can plan hierarchically, is truly creative. They're still far from humans' capabilities and we're still far from achieving that, but we're getting closer at each progress.
Previously, DeepMind has started to interpret AI to the world around it in a similar manner to that of a human baby. This enabled the AI to taught itself to walk and understand the world around it.