RL_lspi_svd

RL SVD feature policy iteration

For 50 iterations or until A_next is identical to A

1)Sample according to policy Yana

K=number of fetures

weights_list=np.zeros(n_actions,K)

2. For each action: 2.1) Crate matrix A, A_next for each action

2.2) Calculte low rank approximation

2.3) Calculate by Woodbery Identity invers e for calculation of weights

Lera

weights_list[action,:]=weight

3. Update for agent our weight

agent.set_weights(weights_list)

4. Sample according to agent policy

We should use flattened env.rander('rdg_array') saved as np.array as observation

Policy is agent.predict(obs) Back to step1

Experiments with Bellman error -> Vlad