# How can we find the value function by solving a system of linear equations without knowing the policy?

An MDP is a Markov Reward Process with decisions, it’s an environment in which all states are Markov. This is what we want to solve. An MDP is a tuple $$(S, A, P, R, \gamma)$$, where $$S$$ is our state space, $$A$$ is a finite set of actions, $$P$$ is the state transition probability function,

$$P_{ss'}^a = \mathbb{P}[S_{t+1} = s' | S_t = s, \hspace{0.1cm}A_t = a] \label{1}\tag{1}$$

and

$$R_s^a = \mathbb{E}[R_{t+1}| S_t =s, A_t = a]$$

and a discount factor $$\gamma$$.

This can be seen as a linear equation in $$|S|$$ unknowns, which is given by,

$$V = R + \gamma PV \hspace{1mm} \label{2}\tag{2}$$

$$V$$ is value of a state vector, $$R$$ is immediate reward vector, $$P$$ is transition probability matrix, where each element at $$(i,j)$$ in $$P$$ is given by, $$P[i][j] = P(i \mid j)$$ i.e., probability that I am in state $$j$$ going to state $$i$$.

As $$P$$ is given, we treat, equation $$\ref{2}$$ as a linear equation in $$V$$. But $$P[i][j] = \sum_a (\pi(a \mid j) \times \mathrm{p}(i \mid j, a) )$$. But, $$\pi (a \mid s)$$ (i.e., probability that I will take action a in state s) is NOT given.

So, how can we frame this problem as the solution to a system of linear equations in \ref{2}, if we only know $$P^a_{ss'}$$ and we do not know $$\pi(a \mid s)$$, which is needed to calculate $$P[i][j]$$?

Your equations all look correct to me.

It is not possible to solve the linear equation for state values in the vector $$V$$ without knowing the policy.

There are ways of working with MDPs, through sampling of actions, state transitions and rewards, where it is possible to estimate value functions without knowing either $$\pi(a|s)$$ or $$P^{a}_{ss'}$$. For instance, Monte Carlo policy evaluation or single-step TD learning can both do this. It is also common to work with $$\pi(a|s)$$ known but $$P^{a}_{ss'}$$ and $$R^{a}_{s}$$ unknown in model-free control algoithms such as Q learning.

However, in your case, you are correct, in order to resolve the simultaneous equations you have presented, you do need to know $$\pi(a|s)$$

This is not as limiting as you might think. You can construct a control method using simultaneous equations, by starting with the policy set to some arbitrary policy. Either a randomly-chosen deterministic policy or the equiprobable policy are reasonable first guesses. Then, after each solution to linear equations, you improve the policy so that each action choice maximises the expected return. This is essentially the policy iteration algorithm but replacing the policy evaluation step with the linear equations method for calculating the values.

• @ Neil Slater. Thanks. I understood. Also one more question I have that is, are we assuming the fact that if there exist a stochastic policy then there exist an optimal deterministic policy? – Abc1729 Feb 16 at 9:55
• @Abc1729 For a well-defined MDP with the Markov property, there will always exist one or more optimal deterministic policies. You do not require a stochastic policy for it to be optimal (but you might want one if your optimisation porcess involves exploring by sampling actions to refine estimates). There are other questions about that on this site already, but if you cannot find them or the answers on them do not help, then feel free to ask a new one – Neil Slater Feb 16 at 11:53
• @Abc1729 if the answer resolved your issue, kindly accept it - see What should I do when someone answers my question? – desertnaut Feb 17 at 14:58