The artificial intelligence division of the search giant is developing artificial intelligence technology that can make its own plans. According to DeepMind, its “Imagination-Augmented Agents” can “imagine” the possible outcome of its actions, and interpret those simulations. They can make the right decision, according to DeepMind, for whatever they plan to achieve.

The AI think and then they learn: Report

According to the researchers, they outperformed baseline agents considerably in a pair of tasks. The AI essentially thinks like humans do and tries out different strategies in their heads (theoretically speaking) so that they are able to learn despite having little real experience. In a blog post, the researchers wrote, “The agents we introduce benefit from an ‘imagination encoder’ – a neural network which learns to extract any information useful for the agent’s future decisions, but ignore that which is not relevant.”

The researchers wrote in the blog post, “Imagining the consequences of your actions before you take them is a powerful tool of human cognition. When placing a glass on the edge of a table, for example, we will likely pause to consider how stable it is and whether it might fall. On the basis of that imagined consequence we might readjust the glass to prevent it from falling and breaking. This form of deliberative reasoning is essentially ‘imagination’, it is a distinctly human ability and is a crucial tool in our everyday lives.”

The researchers have tested the imagination-augmented agents on a spaceship navigation game and the puzzle game ‘Sokoban’. Both the games require reasoning and forward planning.

The AlphaGo program of DeepMind was different: DeepMind

In the blog post, the researchers add that the imagination-augmented agents outperform the imagination-less baselines considerably for both tasks and they “learn with less experience and are able to deal with the imperfections in modelling the environment.”

They add, “Because agents are able to extract more knowledge from internal simulations they can solve tasks more with fewer imagination steps than conventional search methods, like the Monte Carlo tree search.” The AlphaGo program of DeepMind was different and the AI operated in a perfect environment in that program with clearly defined rules that allow outcomes to be “predicted very accurately in almost every circumstance.”

The AlphaGo program of DeepMind beat the best human Go players of the world. The company wants to create computers now that can thrive in imperfect environments, which are complex and where unpredictable problems can arise.