agent_loop in evaluation.py randomizes agent types

[philipjhj]

In the agent_loop in evaluation.py (with a fix for #43 used)

[...]
while episode < EVAL_EPISODES:
        # check if env needs reset
        if env.done:
            print('Episode %d (%.2f)%%' % (episode, (episode / EVAL_EPISODES) * 100.))
            
            # **** fixes #43
            if type(agent.current_agent) == RandomAgent:
                    agent_type = PigChaseEnvironment.AGENT_TYPE_1
                else:
                    agent_type = PigChaseEnvironment.AGENT_TYPE_2
            
            obs = env.reset(agent_type)
            # ****
            while obs is None:
                # this can happen if the episode ended with the first
                # action of the other agent
                print('Warning: received obs == None.')
                obs = env.reset()

            episode += 1

        # select an action
        action = agent.act(obs, reward, agent_done, is_training=True)
        # take a step
        obs, reward, agent_done = env.do(action)

Since agent.act([...],agent_done=True,[...]) resets the current_agent used, it needs to happen before env.reset(agent_type). Otherwise the appearance of the agent is based on the agent type used during the previous episode.

By simply moving the agent.act([...]) call above the if env.done, leads to an action from the previous episode to be given as the first action in the subsequent episode.