def main():
# initiates trainer, actions are these predictions of the agent:
# action 1 -> right box will fall off the edge
# action 0 -> right box will not fall off
trainer = TpgTrainer(actions=[0, 1], teamPopSize=50)
_min, _max, _avg = [], [], [] # hold values for every generation
for gen in range(GENERATIONS): # generation loop
print("Generation: ", gen + 1, "/", GENERATIONS)
curScores = [] # new list per gen
while True: # loop to go through agents
agent = trainer.getNextAgent()
if agent is None:
break # no more agents, so proceed to next gen
# evaluting the agent
score = 0
for i in range(EVALUATIONS):
score += evaluateAgent(agent)
agent.reward(score)
curScores.append(score)
print("Min:", min(curScores), " Max:", max(curScores), " Avg:",
sum(curScores) / len(curScores),
"(out of " + str(EVALUATIONS) + ")\n")
_min.append(min(curScores))
_max.append(max(curScores))
_avg.append(sum(curScores) / len(curScores))
trainer.evolve()
# getting best agent after all the generations
best_agent, best_score = getBestAgent(trainer)
print("Best agent's score:", best_score, "/", EVALUATIONS)
for run in range(FINAL_RUNS):
print("Final run: ", run + 1, "/", FINAL_RUNS, end='\r')
evaluateAgent(best_agent, graphics=True)
# plotting progress over the generations
generations = range(1, GENERATIONS + 1)
axes = plt.gca()
axes.set_ylim([0, EVALUATIONS])
plt.plot(generations, _min, label="min")
plt.plot(generations, _max, label="max")
plt.plot(generations, _avg, label="avg")
plt.xlabel("generation")
plt.ylabel("score")
plt.legend()
plt.show()
for gen in range(100): # generation loop
curScores = [] # new list per gen
# get right env in envQueue
game = gameQueue.pop() # take out last game
print('playing on', game)
env = envs[game]
# re-get games list
if len(gameQueue) == 0:
gameQueue = list(allGames)
random.shuffle(gameQueue)
while True: # loop to go through agents
teamNum = trainer.remainingAgents()
agent = trainer.getNextAgent()
if agent is None:
break # no more agents, so proceed to next gen
# check if agent already has score
if agent.taskDone():
score = agent.getOutcome()
else:
state = env.reset() # get initial state and prep environment
score = 0
valActs = range(env.action_space.n)
for i in range(1000):
act = agent.act(getState(state),
valActs=valActs) # get action from agent