Reinforcement Learning

We studied optimal exploration and developed a good strategy for learning about certain types of unknown environments

Some problems require a computer to control a system, but instead of having full knowledge of the system, the computer is only given partial knowledge of the system and must learn the rest by experience, and instead of giving the computer an explicit goal, it is only told if it is doing well. Some examples of these systems include helicopter control (where the system is penalized for crashing or for taking a long time getting to a target), cross channel marketing (where the system has some control over marketing material in one channel and is rewarded for increasing sales in another channel), and game playing (where the system is rewarded for beating the opponent).

Problems of this type are called reinforcement learning problems, and they are characterized by the computer having to trade off between exploring its environment so it can learn how to get the most reward, and exploiting its environment to get the most reward. Typically if the computer tries only to get the most reward given what it already knows, then it will never find better ways to get reward. However, trying to find better ways to get reward usually involves trying strategies that aren’t very good.

Another challenge of reinforcement learning is that most real world tasks are very complex, and a lot about them is already known. If the computer were to start from scratch, it would take a very long time for it to learn enough to behave in a useful fashion. This is solved by telling the computer as much as possible about what is already known about the task. The way this is done depends on the nature of the task and the information available about it. Partly because of this task dependence, there are a lot of different techniques available for doing reinforcement learning.

If you have a problem that might be solvable using reinforcement learning, we would be glad to talk to you about it, so please contact us today!