Wait what? AI can imagine???
Yeah… As human, we have the ability to think before do an action. It is a powerful tool for human. If we look in to the real world, there are many complex and unpredictable problems arise. As an example, if you keep your iPhone on the edge of the table, you will consider that it might fall or it is safe at there. Using the basis of that imagined value, you will take actions in order to save your iPhone. These types of premeditated reasoning are ‘imagination’, and it is a significant ability of the human race in everyday life of ours.
As per as this situation, in the development of Artificial intelligence, DeepMind which is the world top artificial intelligence research centre of Google developed AI which can ‘imagine’ before perform tasks with out rely on the directions and instructions of human. DeepMind said that “‘Imagining the consequences of your actions before you take them is a powerful tool of human cognition,’ DeepMind said” This AI is called as Imagination-Augmented Agents, or I2As. They use internal ‘imagination-encoder’ which helps them to extract information that might be useful for future decisions, while ignoring anything irrelevant. And with that they can reason about the future while they can construct plans using the results of the imagination and can do deliberate reasoning and can learn different strategies to construct plans, choosing from a broad spectrum of strategies. Researchers said that “They (I2As) learn with less experience and are able to deal with the imperfections in modelling the environment.”
The I2As adapt a number of imagined routes to solve the problem. The encoder also improved the efficiency of the system, which is able to extract additional information from imagination beyond rewards — these routes may contain useful clues even if they do not necessarily result in high reward. The researchers of DeepMind said that “They can learn different strategies to construct plans. They do this by choosing between continuing a current imagined trajectory or restarting from scratch. Alternatively, they can use different imagination models, with different accuracies and computational costs. This offers them a broad spectrum of effective planning strategies, rather than being restricted to a one-size-fits-all approach which might limit adaptability in imperfect environments”. The Imagination-Augmented Agents combines trial-and-error learning with simulation so as to evaluate the most capable trails while sidestepping dead ends.
The specific difference in this I2As is they use noisy data to get learned instead of rely on sanitized information such as a pre-specified. And they create, evaluate and follow through on plans. These AIs create ‘Imagination tress’ which helps them to continue imagining from their last imaginary situation (previously imagined state) formed meanwhile it’s in the last action.
As a testing scenario, the DeepMind researchers gave the Imagination-Augmented Agents, the puzzle game Sokoban and a spaceship navigation game. These games need a previous planning and reasoning. The puzzle game Sokoban is where the player have to move the boxes to get them to storage locations specified/spotted. Boxes may not be pushed into other boxes or walls, and they cannot be pulled. The puzzle is solved when all boxes are at storage locations. Because boxes can only be pushed, many moves are irretrievable.
For these both tasks, the I2As perform better than the AIs with out imagination. The researchers said “For both tasks, the imagination-augmented agents outperform the imagination-less baselines considerably: they learn with less experience and are able to deal with the imperfections in modelling the environment”. Further they add “When we add an additional ‘manager’ component, which helps to construct a plan, the agent learns to solve tasks even more efficiently with fewer steps”. For the other puzzle, the spaceship game also they said that the Imagination-Augmented Agents distinguish between situations where the gravitational pull of its environment is strong or weak, meaning different numbers of these imagination steps are required. For both tasks, the Imagination-Augmented Agents overtake the AIs without imagination and they learn with a low amount of experience and they are well learned to deal with the imperfection in the environment. They said that this Imagination-Augmented Agents managed to solve 85 per cent of the Sokoban levels presented, compared to 60 per cent for a standard model-free agent. DeepMind said that “When it comes to ability to plan and ability for future reasoning, there’s still a lot of work to be done, but this first look is a promising step towards imaginative AI”.