With the plethora of artificially intelligent machines in science fiction, and the general lack of robot companions in reality, many people think that AIs are purely fictional. While we may not have machines able to understand and tell a joke (though those are being worked on) multiple institutions have computers capable of machine learning; a form of AI. Now researchers at MIT have developed a new machine learning framework for reinforcement learning that can complete experiments in a fifth the time of other algorithms.
Reinforcement learning requires two components; an agent and the reward function. The agent is the machine learning what it needs to complete a task and the reward function scores its policies on how effective they are. This kind of algorithm works well for scenarios such as network administration, but can become cumbersome as millions of combinations of actions may have to be considered. The researchers have gotten around that though with their reinforcement learning and Python, or RLPy, framework. This software allows the computer to construct a tree of every possible combination and work its way from the simplest to most complex combinations. What makes it more efficient though is that once the computer realizes that the descendants of a specific combination always return the same result, because of that combination, it will stop testing descendants of that combination.
The RLPy software has already be released online along with its source code, so anyone can use and develop it further. For the time being the researchers are using it to teach an autonomous vehicle how to drive.