bellman_equation.md

tags	python,numpy,neural-network,reinforcement learning
mathjax	true

The Bellman Equation

The Bellman Equation defines the value $$V^\pi(s)$$ of a given state as the sum of

• the immediate reward $$r(s, a)$$ received when executing action $$a$$ in state $$s$$ following policy $$\pi$$
• the value $$V^\pi(s')$$ of state $$s'$$ reached by executing the policy step

This gives the following bootstrap state value evaluation rule:

$$V^\pi(s) = r(s, a^\pi) + V^\pi(s')$$

for deterministic environments.

For stochastic environments, in which the same action $$a$$ executed in the same state $$s$$ may result in different next states $$s'$$, the transition probabilities $$p$$ have to be taken into account.

The next state value $$V^\pi(s')$$ becomes the sum over all next states, that can be reached from state $$s$$ executing policy action $$a^\pi$$, weighted with the related state transition probabilities $$p$$:

$$V^\pi(s) = r(s, a^\pi) + \sum_{s'} p(s'|s, a^\pi) V^\pi(s')$$

Because the value of each state depends on the value of the related next state, it becomes the sum of future rewards collected while executing a given policy.

To prevent infinite state values, the sum of future rewards often gets discounted by a factor $$\gamma$$ smaller than 1:

$$V^\pi(s) = r(s, a^\pi, s') + \gamma \sum_{s'} p(s'|s, a^\pi) V^\pi(s')$$

The Q state-action value (state-action pair quality) related to state $$s$$ and action $$a$$ is defined as:

$$Q(s, a^\pi) = r(s, a^\pi, s') + \gamma \sum_{s'} p(s'|s, a^\pi) \max_{a'} Q(s', a')$$

which can be transformed into a $$Q$$ value Temporal Difference (TD) Learning rule with:

$$G^\pi = r(s, a^\pi, s') + \gamma \sum_{s'} p(s'|s, a^\pi) \max_{a'} Q(s', a')$$

and the learning rate $$\alpha$$, we get the following $$Q$$ value update rule:

$$Q_{t+1}(s, a^\pi) = (1 - \alpha) Q_{t}(s, a^\pi) + \alpha G^\pi_{t}$$

The state transition probabilities are a property of the environment and must not be explicitly implemented by a TD learning algorithm. The transition probabilities are implicitly modelled (emerge) by doing a lot of averaging TD steps with the same state action pairs and probably different next states.

TD Learning Algorithms

The following temporal difference learning algorithms are based on the Bellman Equation:

• Q-Learning Algorithm
• SARSA Algorithm
• Expected SARSA Algorithm
• Dyna-Q Algorithm

Please see the Gridworld Examples for details.