Open droftware opened 7 years ago
Observation 1: Target random variable is clearly being over shadowed by Obstruction and Visibility.
Consider the last direction where Obst:3, Vis:1 and Target:F , the corresponding posterior probability is 0.232 . Now consider the direction where Target:T, the values are 0.006 (where Obs:1, Vis:2) and 0.0003(where Obs:0, Vis:0). These values are no match to the posterior direction values where Obs or Vis is >= 2.
All random vars set ('Target:', ['F', 'F', 'F', 'F', 'F', 'F', 'F', 'F', 'F', 'F', 'F', 'T', 'T', 'F', 'F', 'F', 'F']) ('Danger:', ['0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0']) ('Obstruction:', ['3', '3', '0', '3', '2', '2', '2', '3', '3', '0', '0', '0', '1', '1', '1', '0', '3']) ('Visibility:', ['0', '0', '0', '0', '1', '1', '1', '0', '0', '0', '0', '0', '2', '2', '2', '0', '1']) ('Hider:', ['0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0']) ('Seeker:', ['0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0']) ('Blockage:', ['0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0', '0']) [0.15699839127296644, 0.15699839127296644, 0.00025993939072049556, 0.15699839127296644, 0.0046280652395040891, 0.0046280652395040891, 0.0046280652395040891, 0.15699839127296644, 0.1140213695038836, 7.6297964582254014e-05, 1.986822018648759e-05, 0.00037123910581614023, 0.0062351650998350387, 0.0013643480494011468, 0.0028666336809768404, 0.00025975764832106074, 0.23264762052589899]
One of the ways to circumvent the problem caused due to the above Observation 1 is to have three major behavioural types for the hider. Each behavioural type will be defined by its own customised bayesian network having a subset of the existing random variables.
The three behavioural types will be:
Curious This will enable the agent to be more explorative and braver. Its random variables will be Visibility,Danger,Hider,Seeker,Blockage. Priors for visibility will be increased.
Scared This will enable the agent to be extra cautious and suspicious. Its random variable will be Obstruction, Danger,Hider, Seeker, Blockage. Priors for Obstruction will be increased.
Mobile This will enable the agent to be more mobile and allows it to traverse to a different region from its current region. Its random variables will include Target, Danger, Hider, Seeker, Blockage. Priors for Target will be increased.
Sidenote: Changing the above behavioural types of the agents as well as giving instructions regarding movement to other regions will be handled by a higher level decision making layer. This higher level layer will be controlled by a few commanding agents and not all. Under each commanding agent there will be some agents assigned and it will be the duty of the commanding agent to assign and change behaviours as well as give movement related instructions depending on the their higher level view of the game.
Postponing this issue The exact probability vales for the prior and conditional distributions can only be decided by analysing the simulation when the higher level decision making layers are coded and integrate with the lower level bayesian nets.
Though the path planning and bayesian controller have been integrated, the agent is still not respecting the path i.e. it's not traversing to its designated goal position due to intermingling of various probabilities of many random variables such as visibility, obstruction etc.
The bayesian network as well the probability tables of the associated random variables need to be tweaked so as to enable the agent in respecting its goal destination of its formulated path.