REnforce

fn main() {
	// Create the environment
	let mut env = ExampleEnvironment::new();
	
	// Here, the agent will use linear value approximators for each action in a given state
	// The agent will select actions based on how high of a value it assigns them
	let q_func = QLinear::default(&env.action_space());
	
	// Creates an epsilon greedy Q-agent
	// Agent will use softmax to act randomly 5% of the time
	let mut agent = EGreedyQAgent::new(q_func, env.action_space(), 0.05, Softmax::default());
	
	// Here, we use Q-learning to train the agent
	// By default the discount factor is 0.95,
	//            the learning rate is 0.1,
	//            the trainer trains for 100 episodes when called train
	// We set the learning rate (alpha) to 0.9
	let trainer = QLearner::default(&env.action_space()).alpha(0.9);

	// Magic happens
	trainer.train(&mut agent, &mut env);

	// Simulate one episode of the environment to see what the agent learned
	let mut obs = env.reset();
	while !obs.done {
		env.render();

		let action = agent.get_action(obs.state);
		obs = env.step(action);
	}
	env.render();
}

Contributing

If you see something that could be done better, or have an idea for a feature that you want to add, then fork and sumbit a PR and/or create an issue for it. If you want to contribute, but you're not sure where to start, feel free to email me with any questions.

Progress

A lot remains to be done, but the following reinforcement learning algorithms have been implemented* thus far...

• Q-learning
• SARSA
• Cross Entropy
• Vanilla Policy Gradients
• Natural Evolution Strategies

* - implementations possibly flawed, but you have to start somewhere