Add actor loss coefficient hyperparameter to A2C

Author: Omegastick

example/gym_client.cpp

@@ -131,7 +131,7 @@ int main(int argc, char *argv[])
     std::unique_ptr<Algorithm> algo;
     if (algorithm == "A2C")
     {
-        algo = std::make_unique<A2C>(policy, value_loss_coef, entropy_coef, learning_rate);
+        algo = std::make_unique<A2C>(policy, actor_loss_coef, value_loss_coef, entropy_coef, learning_rate);
     }
     else if (algorithm == "PPO")
     {

example/lib/libzmq

@@ -16,11 +16,12 @@ class A2C : public Algorithm
 {
   private:
     Policy &policy;
-    float value_loss_coef, entropy_coef, max_grad_norm, original_learning_rate;
+    float actor_loss_coef, value_loss_coef, entropy_coef, max_grad_norm, original_learning_rate;
     std::unique_ptr<torch::optim::RMSprop> optimizer;
 
   public:
     A2C(Policy &policy,
+        float actor_loss_coef,
         float value_loss_coef,
         float entropy_coef,
         float learning_rate,

include/cpprl/algorithms/a2c.h

@@ -14,13 +14,15 @@
 namespace cpprl
 {
 A2C::A2C(Policy &policy,
+         float actor_loss_coef,
          float value_loss_coef,
          float entropy_coef,
          float learning_rate,
          float epsilon,
          float alpha,
          float max_grad_norm)
     : policy(policy),
+      actor_loss_coef(actor_loss_coef),
       value_loss_coef(value_loss_coef),
       entropy_coef(entropy_coef),
       max_grad_norm(max_grad_norm),
@@ -90,7 +92,7 @@ TEST_CASE("A2C")
         ActionSpace space{"Discrete", {2}};
         Policy policy(space, base);
         RolloutStorage storage(5, 2, {1}, space, 5, torch::kCPU);
-        A2C a2c(policy, 0.5, 1e-3, 0.001);
+        A2C a2c(policy, 1, 0.5, 1e-3, 0.001);
 
         // The reward is the action
         auto pre_game_probs = policy->get_probs(
@@ -160,7 +162,7 @@ TEST_CASE("A2C")
         ActionSpace space{"Discrete", {2}};
         Policy policy(space, base);
         RolloutStorage storage(5, 2, {1}, space, 5, torch::kCPU);
-        A2C a2c(policy, 0.5, 1e-7, 0.0001);
+        A2C a2c(policy, 1, 0.5, 1e-7, 0.0001);
 
         // The game is: If the action matches the input, give a reward of 1, otherwise -1
         auto pre_game_probs = policy->get_probs(