25th LSI Design Contests-in Okinawa  Design Specification - 4

4. Maze exploration using reinforcement learning

The following file is this example in Matlab.

Zip file(m file):DQN_sample_e.zip
How to use: Run sw_Q _ Learning.m from the DQN_sample folder

As an example using reinforcement learning, we deal with the 3 × 3 maze search problem shown in Fig. 3.


Fig 3 : 3×3 maze

Agents: People
State: Where in S1 ~ S9 the agent is

Action: Move in the direction of「→」,「↑」,「←」,「↓」
            S5, S7, S8: Negative reward (demon)
            S9: Positive reward (money)
            Else: No reward

In the case of this maze search problem, the purpose of deep reinforcement learning is to get maximum reward (money) at the time of goal (When we arrived at S 9,).
→ Do not go through a trout with demons.

<<Back                 Next>>