It provides code for creating agent entities that are capable of interacting with an environment (provided by user) in a manner that humans would call a reasonable behaviour. But what does it actually mean? Well, that means that our environment (bunch of states and transition rules) has to be
- observed (fully or partially) at any given moment of time,
- acted on (in a finite or infinite number of ways) at any given moment of time. That also means that we are generating a set of rules which ouput actions when provided with observations. Now wait a minute, if we are to have a computer program that interacts with environment (instead of us doing it manually) we want those actions to be meaningful. We don't want them to be stupid actions, don't we? Anyone can output stupid actions. And how would we ensure that those actions are meaningful? Well, for that reason we better have our environment to carry a
- fitness function, usually implemented via reward. It provides user with some feedback. Big rewards mean we are doing okay, small rewards mean we have to try harder (or rather our agent has to try harder). Good thing mathematicians came up with algorithms that solve exactly this problem. *Reinforcement learning: agent interacts with an environment and maximizes expected reward over time. Back to code.
we want to have mechanisms for training agents, evaluating agent's performance and saving agent's behavior for further use. This is what this code helps with. It is also designed to allow carrying out experiments: comparing different agents in various ways. If you want to use this code in your project, checking out examples is a good start. DeepRL doesn't have any documentation and won't have any time soon.
Better install these python packages or it won't work:
- pytorch
- gym[Box2D]
- numpy
- pandas
- matplotlib