void pollAlternativeAction(IAgent agent)
{
// This will be a negative (bad) reward.
// We need to handle this by polling other agents to see if they
// could have avoided it.
if (TeachParadigm != TeachingParadigm.EveryonePolling &&
TeachParadigm != TeachingParadigm.EveryoneRewardsAndPolling &&
TeachParadigm != TeachingParadigm.SubculturePolling &&
TeachParadigm != TeachingParadigm.SubcultureRewardsAndPolling)
{
return;
}
// If we are using EveryonePolling, the agent can ask the entire population.
// If we are using SubculturePolling, the agent can only ask agents in its subculture.
var available = TeachParadigm == TeachingParadigm.EveryonePolling ? _agents.Where((a, i) => agent.Id != i)
: _agents.Where((a, i) => _agentGroups[agent.Id] == _agentGroups[i] && i != agent.Id);
// The teacher is the highest fitness agent available to the eaten agent.
var teacher = (SocialAgent)available.OrderByDescending(a => a.Fitness).FirstOrDefault();
// Get the corrected moves for every action in the eaten agent's memory
LinkedList <StateActionReward> badTrajectory = ((SocialAgent)agent).Memory;
LinkedList <StateActionReward> correctTrajectory = new LinkedList <StateActionReward>();
foreach (var bad in badTrajectory)
{
double[] state = bad.State;
double[] action = new double[bad.Action.Length];
teacher.activateNetworkWithoutMemory(bad.State).CopyTo(action, 0, action.Length);
StateActionReward good = new StateActionReward(state, action, 0);
correctTrajectory.AddLast(good);
}
// Train the agent with the correct trajectory according to the teacher.
TeachAgent(agent, correctTrajectory);
}
