Route-Gym

An openai gym that allows agent to solve shortest and longest route problems.

These simples are simple for humans, and can be computed easily by hand using Dijkstra's algorithm and its variants like A* (computation time permitting). As we work towards AGI, it is of my opinion that our complicated general-purpose algorithms should be asked to solve these simple problems. If convergence is slow, or if the supposedly "general" algorithm can't solve it easily, it is safe to say that that algorithm should be reworked.

Quickstart

env = ShortestRouteEnv(nx.frucht_graph(), 0, 5, random_weights=(1,10))
env.render()    # optionnal
done = False
# you might want to give the adjacency matrix to the policy
policy = ?
rew = 0
position = origin
while not done:
    action = policy.predict(position)
    position, reward, done, _ = env.step(action)
    end.render()    # optionnal
    rew += reward
print("Final reward:", rew)
print("Dijkstra's reward:", env.graph.dijkstra_rew)

What is provided in this gym?

The environment:

The two environments are based neton OpenAI's gym.

The environment can be called using routegym.env.ShortestRouteEnv or the equivalent for the longest route version.

The environments have a render function you can use to display the environment's state. In it, the blue path on the graph's arcs represents Dijkstra's path. This only works for ShortestRouteEnv.

The environments receive a networkx graph, an origin, a goal, and random weight boundaries (if need be) as part of their constructor.

You can also set the make_horizon flag to True to transform the graph into a finite-horizon problem. Be warned that this should only be used on smaller graphs: this generates a tree out of all the possible paths the agent can take from origin to goal and merges them into a single graph. Needless to say, the big O of this thing is enormous! This flag should only be used for toy examples.

The Graph class

The environments use a custom graph class as a backend. A typical user should not need to interact with this class.

But this class does calculate the Dijkstra solution for the problem. If you want to compare your algorithm's performance to Dijkstra's, you can use env.graph.dijkstra_path to get Dijkstra's path, or env.graph.dijkstra_rew to get Dijkstra's reward.

The validate script

This is an internal test, but it is provided as a courtesy to the users. You can get inspiration from that script, either as a tutorial on how to use this package, or as a way to generate many environments, etc.

You can view it in routegym.validate.py.