The ddpg_numpy_only from kcg2015

ddpg_numpy_only's Introduction

Overview

In this project, we implement Deep Deterministic Policy Gradient (DDPG) from scratch (using Numpy only), without using DL framework such as Tensorflow.

DDPG

Key steps in DDPG

In the following, the key steps and their corresponding code snippets are listed. The DDPG algorithm is implemented in ddpg_numpy.py

Select action a_t according to current policy and exploration noise

a_t = actor.predict(np.reshape(s_t,(1,3)), ACTION_BOUND, target=False)+1./(1.+i+j)

Execute action a_t and observe reward r_t and observe new state s_{t+1}

s_t_1, r_t, done, info = env.step(a_t[0])

Create and sample from replay buffer

Set y_i according to the following equation:

y=np.zeros((len(batch), action_dim))
a_tgt=actor.predict(states_t_1, ACTION_BOUND, target=True)
Q_tgt = critic.predict(states_t_1, a_tgt,target=True)
for i in range(len(batch)):
    if dones[i]:
        y[i] = rewards[i]
    else:
        y[i] = rewards[i] + GAMMA*Q_tgt[i]

Update the critic network by the loss function

loss += critic.train(states_t, actions, y)

Update the actor policy using the sampled policy gradient:

which needs the input of

dQ_da = critic.evaluate_action_gradient(states_t,a_for_dQ_da)

which in turn relies on $a=\mu(s_i)$:

a_for_dQ_da=actor.predict(states_t, ACTION_BOUND, target=False)

Finally, the following code implements the actor policy update:

actor.train(states_t, dQ_da, ACTION_BOUND)

Update target networks

actor.train_target(TAU)
critic.train_target(TAU)

Actor (policy) Network

The actor network is implemented in actor_net.py .

Critic (value) Network

The critic network is implemented in critic_net.py . We note that we follow the implementation as mention in DDPG paper. The following sketch shows the architecture of the critic network.